MultiBody coverage work

This is the working evidence ledger for multibody-library-coverage, based on main commit 97eb3ab74b3e11264ab2000437eb47df1a57214d. Work is in progress; complete MultiBody support has not been established.

Runtime indexing prerequisites for variable coordinate projections

The coordinate-switching investigation found two general classification bugs before a variable projection could execute. For der(x) = values[k], Solve’s direct-derivative query attempted to resolve the value-side index at compile time. The existing runtime gather was therefore unreachable. The query now first proves that its expression contains a derivative. With the equation reversed, DAE partitioning mistook the Integer index for a discrete assignment target. Assignment-side classification now follows the selected value through references, indices, unary wrappers, arrays, and tuples, excluding the index expressions. Discrete targets and malformed discrete solved forms retain their checks. These fixes follow MLS §10.5 / ARR-005, ARR-024 and the Appendix B partition contract; they add no Modelica exception, IR vocabulary, or public API.

runtime_index_derivative reproduces the original lowering failure and verifies both orientations of a tensor derivative equation whose index changes at an event. Its analytical integrals pass; an out-of-bounds runtime index still fails during execution instead of producing a successful trace. The ordinary worker/OMC comparison in variable-projection-prototype-1 reports seven compared channels, all high. All 271 DAE library tests, 144 Solve library tests, and 677 core tests pass, along with affected-phase Clippy, core-test Clippy, and workspace formatting. A test-only nesting warning was corrected and the focused tests and Clippy rerun.

Tier 1 runtime-index-msl-focused-1 retains three compared/high models; runtime-index-canary-1 retains nine compared/high and eleven existing failures. Both have zero phase/band changes and zero skipped/missing traces against their gyro-matrix-refinement predecessors. No new Tier 2 claim or baseline promotion is made. runtime-index-derivative-evidence-1.json binds the exact test, worker, OMC, and comparator evidence under .git/multibody-campaign/rolling-wheel.

The local SelectedCircle diagnostic now executes the variable-index projection before the turning point, with all four physical position/velocity errors below 8.0e-9 against the analytical solution. This is not automatic coordinate switching. The authored SwitchedCircle probe still fails during event projection; OMC cannot build either projection probe because its backend overflows its stack. Neither probe earns OMC parity or coverage credit. The original CircleChart wrong-branch counterexample and GyroscopicEffects failure remain open. The remaining compiler/runtime work is a construction-issued coordinate family and an atomic physical-state-preserving FMI transition.

Static coordinate folds expose a circular-motion counterexample

After the sensitivity correction in 14500e1264754060dffe3099b7e9aaf88b57deff, GyroscopicEffects reaches a different failure at t=0.25309309043358214. The failed primal block has 60 coordinates and a four-quaternion tearing plan; this failure does not bypass tearing. Its condition number grows from about 3.8e9 to 1.4e10 across the three captured attempts. The selected position components of bodyCylinder4 cannot be reconstructed: the source-bound parameters give fixed radius squared 0.037144660940672626, while the trial r_0[2], r_0[3] require r_0[1]^2 = -5.846693632395117e-7 in the first attempt. All three attempts are geometrically infeasible. Increasing Newton iterations or relaxing residual tolerances cannot recover a real solution. OMC’s original trace crosses r_0[1]=0 between 0.25 and 0.26 seconds using dynamic quaternion state sets. gyro-primal-geometry-proof-1.json retains the parameter-derived calculation and the matrices/coordinates it diagnoses.

The reduced crates/rumoca/tests/fixtures/index_reduction/CircleChart.mo isolates the problem: der(q)=v, der(v)=lambda*q, q*q=1, initialized at q={1,0}, v={0,1}. The structural DAE selects the aggregate {q[2], $formal_derivative.1.q[2]}. Its physical solution is q={cos(time),sin(time)}, v={-sin(time),cos(time)}, lambda=-1. OMC generates independent dynamic position and velocity state sets and completes a full revolution; maximum absolute analytical error is 1.26e-6. Rumoca fails the full revolution at t=4.712251253211418, with the selected position outside the unit circle.

More seriously, the identical equations with stop time 3 produce sim_ok but reflect onto the wrong branch after the first quarter turn. The ordinary trace comparator reports five compared channels: three high, two deviating, zero minor/severe. Maximum absolute errors are 1.98 in q[1] and 2.00 in v[1]; the initial states agree. The focused analytical Rust regression fails at t=1.58, where q[1]=+0.00920313296 instead of -0.00920354327. This is an open actionable refinement counterexample, outside the fixed MSL cohort. The latest cohort counts remain historical evidence; they do not close this defect. Unrelated capability work, merges, and releases remain blocked by SPEC_0033 until it is resolved.

circle-chart-1 retains the worker requests, all IR stages, complete traces, OMC-generated C/state sets, analytical comparisons, and standard comparator output. circle-chart-regression-red-1.patch and its test log retain the red regression for activation with the fix; no ignored or knowingly failing test is added to the passing suite. All temporary capture instrumentation is removed. gyro-primal-coordinate-evidence-1.json binds these artifacts below .git/multibody-campaign/rolling-wheel. SPEC_0053 §2a records the required dynamic-coordinate contract. It remains unimplemented: basis changes must use compiler-issued candidates, preserve the full physical point and its tangent, and reset the integrator through FMI Event Mode before the current chart folds. No new model or coverage credit is claimed.

Refine algebraic sensitivity solves without weakening certification

GyroscopicEffects fails at t=0.25041127485935627 in a 258-coordinate algebraic sensitivity block. The captured matrix and original directional AD agree on residual row 1479: -1.0296428887592066e-10 against tolerance 1e-10. Although the diagnostic’s matched target is z_a[3], this row actually relates axial torque to 0.015118914645400883 * z_a[1]. Zero torque requires zero axial acceleration; the single LU solve instead introduces 6.810296326875687e-9 into that acceleration direction. The matrix condition number is about 10,110. OMC’s generated six-variable acceleration solve and its near-zero axial trace agree with the source equation. This evidence identifies numerical solve error, not a missing source equation or justification to relax tolerances.

The sensitivity projection now applies at most two matrix-residual correction sweeps when the original global JVP check fails. Each block reuses its same-point matrix/factorization, and downstream right-hand sides see corrected predecessor directions. Acceptance still requires a fresh original JVP and scales from the corrected direction at the original tolerance (SPEC_0038 numerical projection). The correction never fits an inconsistent combined JVP; failure still restores all unknown seeds. Already passing directions take the unchanged fast path.

An analytical free-rotor/load regression fails before the change and passes afterward, including both direction signs and a downstream dependency. Replay of the exact captured original-model point also changes from rejection to success. All 504 solver tests and 675 core tests pass, as do affected-crate Clippy and workspace formatting. Existing inconsistent-JVP and overflow rollback tests remain green. No compiler IR or public API changes.

Tier 1 gyro-matrix-refinement-msl-focused-1 keeps three compared/high models and all 2,633 channels high; gyro-matrix-refinement-canary-1 keeps nine compared/high and eleven existing failures. Both have zero phase/band changes and zero skipped/missing traces against the preceding zero-cost-assignment runs. GyroscopicEffects still fails, now in primal coordinate convergence with residual -1.381412e-6 on the row matched to bodyCylinder3.body.Q[4]. Its current static basis uses position/orientation entries, while OMC uses two dynamic three-of-four quaternion state sets. Whether that difference causes this remaining failure requires a new captured-point analysis. No model gains passing credit from this fix; no new Tier 2 claim or baseline promotion is made. gyro-matrix-refinement-evidence-1.json binds the capture, replay, OMC artifacts, test logs, and Tier 1 deltas below .git/multibody-campaign/rolling-wheel.

Full sweep recovers three MultiBody models without losing high results

The complete multibody-zero-cost-assignment-full run at 9adb545e952bd46431367b9e75b83444bcf71e1d passes the current MSL quality gate: 169/566 high (29.86%), 169 compared, 21 exclusions, no missing traces, and 190 raw completions. MultiBody improves from 20/42 to 23/42 high. RevoluteConstraint, SphericalConstraint, and Fourbar1 recover from Solve timeouts. No previously high model is lost; these are the only band changes. DC_CompareCharacteristics returns from timeout to its structural refusal and earns no passing credit. Stage counts are 566 parsed, 490 flattened, 288 DAE compiled, and 264 Solve compiled. The run uses eleven workers and unchanged budgets, tolerances, target scope, and reviewed exclusions.

The channel tally needs a distinction from the model-level band: all 24,173 compared initial values are high. Trajectories have 24,127 high channels, 46 minor channels, and zero deviating/severe channels. The same 46 minor channels occur in the preceding full run, across seven electrical/magnetic models. Earlier ledger wording conflated the initial-value count with the trajectory count; the two recent entries below now state that distinction. All compared MultiBody trajectory channels are high. Excluded models remain unproved and receive no passing credit.

zero-cost-assignment-full-delta-1.json retains the complete model/phase delta; zero-cost-assignment-full-audit-1.log and zero-cost-assignment-evidence-2.json bind the full-run artifacts and the preceding focused/perf evidence below .git/multibody-campaign/rolling-wheel. The tracked tree was clean at launch; the foreign coordination file accounts for the dirty-worktree marker. No baseline was promoted. This is a full MSL sweep, not combined verify full; release gates and a PR remain pending complete MultiBody coverage.

The next concrete numerical failure is GyroscopicEffects: its focused run compiles in 6.251 seconds and initializes, then rejects sensitivity residual 1479 (bodyCylinder2.body.z_a[3]) at -1.029643e-10 against the unchanged 1e-10 tolerance. The full run reproduces that refusal. Inspection finds one LU solve per sensitivity block followed by a global AD residual check, without residual correction. That suggests investigating linear-solve error versus cancellation in the differentiated equations; it does not justify relaxing the check or claim that refinement will fix the model. The failing matrix, source equations, and OMC reconstruction still need a direct comparison.

Match zero-cost assignment paths before searching

The proposed source-dimension shortcut is rejected. It bypassed ordinary reduction facts needed by derivative aliases, selected tensor components, and assertions. The first broad core run had thirteen failures; a direct-state candidate filter still left nine. Neither implementation nor its proposed spec changes remain. source-formal-profile-rejected-1.patch and the two failed core logs preserve the experiment below .git/multibody-campaign/rolling-wheel.

The retained change addresses the measured assignment cost without bypassing preparation. Differential-signature matching previously ran a heap-based shortest-path search for every matched pair. An available source/row/column/sink path whose three reduced costs are zero is already shortest under feasible potentials. Reversing those zero-cost edges preserves the potentials; matching all such direct paths can precede the existing weighted augmenting searches. Alternating paths still reassign earlier pairs, and the independent perfect matching, dual-inequality, matched-equality, and objective checks required by SPEC_0007 / STRUCT-T07 remain unchanged. This changes no Modelica semantics, tensor owners, initialization requirements, or numerical tolerances.

The actual-worker zero-cost-assignment-revolute-solve-profile-1 records heap pop plus weighted-matching self samples falling from 6.41% to 0.84% against refresh-source-index-revolute-solve-profile-1. Solve lowering measures 8.001 to 6.880 seconds. These are single observations under variable host load, not a timing guarantee. Flat, DAE, structural DAE, Solve IR, and the complete simulation trace are byte-identical. The native probe measures formal construction falling from 0.622 to 0.099 seconds; ordinary preparation remains about 3.93 seconds. zero-cost-assignment-profile-delta-1.json retains hashes, timings, and exact worker identities. Neither diagnostic earns coverage credit.

All 219 structural library tests and 675 core tests pass, as do affected-crate Clippy with all targets/features and workspace formatting. Matching agrees with independent exhaustive enumeration for every 3-by-3 signature. New tests exercise reassignment of an early direct match and certify 4,096 independent derivative pairs with no heap searches.

The unchanged seven-model target list in zero-cost-assignment-msl-focused-1 recovers RevoluteConstraint and SphericalConstraint from Solve timeouts to high parity. HeatLosses remains high: three compared, all 2,633 trajectory channels and initial values high, zero skipped/missing. GyroscopicEffects now reaches its projection-sensitivity rejection after compilation; it still fails. UniversalConstraint still times out, while PointGravity and RollingWheelSetDriving retain their numerical/selection failures. The fixed zero-cost-assignment-canary-1 remains nine compared/high, eleven existing failures, no skipped/missing traces, and zero phase/band delta. These focused results do not update the full-cohort count. No baseline was promoted.

Remove repeated refresh-output scans; compilation timeouts remain open

An actual-worker perf recording of RevoluteConstraint locates repeated whole-program output scans in refresh-owner validation. Each row previously rescanned preceding programs and materialized the complete output-index vector to resolve one canonical output. The checker now builds one borrowed lookup per immutable source block, preserving dense local relocation, sparse logical indices, program output bounds, tensor output cursors, and every assignment certificate check. It adds no serialized metadata or numerical shortcut. This is a read-only construction query under SPEC_0029 §3 and SPEC_0036 / SPEC_0043 §6a; source equations and MLS semantics do not change.

In matched single-worker diagnostics, the output-index routine drops from 4.15% to 0.01% of Solve CPU samples. Measured lowering decreases from 8.281 to 8.001 seconds; this is one observation under variable host load, not a timing guarantee. Flat, DAE, structural DAE, Solve wire, and complete simulation trace are all byte-identical. The exact before/after executable hashes, perf records, requests, phase timings, and artifact hashes are retained in initial-geometry-revolute-solve-profile-1, refresh-source-index-revolute-solve-profile-1, and refresh-source-index-profile-delta-1.json below the evidence directory.

All 325 Solve IR tests and 675 core tests pass, as do affected-crate Clippy and workspace formatting. New lookup tests exercise a million-element compact tensor domain, dense/sparse outputs, multi-output programs, empty blocks, out-of-range identities, and output-range overflow. Existing forged-owner and wire-replay tests remain unchanged and green.

The same seven-model focused run refresh-source-index-focused-1 retains HeatLosses high on 798 channels. Four models time out in Solve, including SphericalConstraint, which completed the previous focused run. PointGravity and RollingWheelSetDriving retain their numerical/selection failures. The optimization therefore does not close the compilation-timeout regression. The fixed canary refresh-source-index-canary-1 remains nine compared/high, eleven existing failures, no skipped/missing traces, and zero phase/band delta. The subsequent complete run multibody-refresh-source-index-full at 955d33e905b1bc9e699936b3f648f0b2810131f4 passes the current quality gate: 166/566 high (29.33%), 166 compared, 21 exclusions, no missing traces, 187 raw completions, and 20/42 MultiBody high. All 21,484 initial values are high; trajectories have 21,438 high and 46 minor channels, with zero deviations. SphericalConstraint’s Solve timeout is the only band loss against multibody-initial-geometry-full; its failure remains open. DC_CompareCharacteristics changes from structural rejection to timeout, and Inverse_sh_TX returns to Flatten rejection; neither was previously high. refresh-source-index-full-delta-1.json retains the per-model comparison. No baseline was promoted, and this MSL run is not a combined verify full.

The public-API diagnostic revolute-structural-cost-1 separates the next cost. Ordinary preparation takes 3.946 seconds, traverses a 16-round direct pass plus holonomic reduction, and returns eight state scalars and four manifold roots. Source formal construction takes 0.622 seconds and proves formal dimension four. Complete lowering takes 7.773 seconds and produces four state scalars. state_selection::prepare discards the ordinary prepared system and constructs that final candidate from the original source. These are diagnostic timings, not coverage evidence. Avoiding the discarded work requires a source-bound applicability decision that preserves initialization, state preferences, structural matching, and numerical checks; bypassing those obligations is not an optimization. The diagnostic source, executable, and log are retained below the evidence directory. refresh-source-index-evidence-1.json binds the checks, profiles, diagnostics, full-run artifacts, and worker hashes.

Preserve stated initial geometry when selecting independent states

The complete run multibody-initial-geometry-full at 9708e306d9b8c6e4bb248155f701ea80a7ddd5d8 passes the MSL quality gate with 188 raw completions, 167/566 compared/high models (29.51%), 21 exclusions, no missing traces, and 21/42 MultiBody high. All 22,401 initial values are high; trajectories have 22,355 high and 46 minor channels, with zero deviations. The tracked worktree was clean at launch; the untracked coordination file accounts for the comparator’s dirty-worktree flag.

HeatLosses gains high parity on all 798 shared trajectory channels and initial values, but RevoluteConstraint loses its prior high result to the ordinary ten-second Solve compilation budget. These are the only band changes from multibody-reference-boundary-full at eacdb83d900a56576b49ecaf2c46f0a06df617c6; the total is unchanged. The phase delta also moves Media.Examples.SolveOneNonlinearEquation.Inverse_sh_TX from Flatten rejection to DAE rejection; it earns no passing credit.

The HeatLosses source fixes body1.r_0 to {0.3,-0.2,0}. Its frame-position alias has an unfixed zero start; the damper length also has an unfixed zero start. Automatic selection previously ignored the exact initial-value transfer when constructing its trial point and treated both source-state guesses with the same priority. It selected damper1.s in place of the Cartesian body1.frame_a.r_0[1]. At the actual initial geometry the damper is vertical: its length has zero first derivative with respect to that Cartesian coordinate. The length therefore cannot locally replace that coordinate. Initialization then reached a non-finite damper-direction projection.

The before/after checked DAE maps both contain 37 independent coordinates. Their only coordinate-set difference is damper1.s → body1.frame_a.r_0[1], matching OMC’s Cartesian state choice. Source Flat and DAE are unchanged. The root fix carries the structural owner’s exact, source-bound initial-value transfers into formal analysis, uses them as trial guesses, and prefers stated initial coordinates within each existing StateSelect/declared-state preference class. Required and forbidden states, full-rank checks, tensor owners, original initial equations, and subsequent numerical reconstruction remain authoritative. Trial values never become runtime seeds or a claim that initialization has been solved.

The independent InitialRadius reproduction combines a Cartesian oscillator, a redundant radius, and a separately constrained rotating vector. Without the constraint, ordinary reduction succeeds; with it, automatic selection fails before the fix. The regression independent_state_basis_respects_fixed_initial_geometry now checks all six observables against their sine/cosine solutions. OMC’s same-source C simulation agrees with those solutions within 2.17e-6 at unchanged 1e-6 tolerance. The generated OMC initialization assigns p[1] and p[2] as Cartesian states.

Holding stated trial values fixed was tested and rejected: it prevented a parameter-dependent fixed start from participating in simultaneous initialization and moved four rejection cases into the wrong failure path. Those failed logs remain evidence. The final implementation has no fixed-column correction or new numerical linear-algebra API. All 675 core tests, 144 Solve and 217 structural library tests, six spec checks, affected-crate Clippy, and workspace formatting pass.

The identical seven-model control/candidate target list is initial-geometry-targets-1.json. The archived control workers reproduce HeatLosses’ original NaN failure; all seven controls fail, including four Solve timeouts, so control parity is unmeasured, not zero. The candidate compares HeatLosses and SphericalConstraint, both high (798/917 channels); UniversalConstraint, RevoluteConstraint, and GyroscopicEffects time out, while PointGravity and RollingWheelSetDriving retain their original failures. A generic band-delta script correctly rejects the missing control band table; initial-geometry-focused-1-phase-delta.json records the phase evidence without inventing old bands. The fixed canary initial-geometry-canary-1 remains nine compared/high, 175 high initial channels, eleven existing failed models, no skipped/missing traces, and zero phase/band delta.

Before/after IR is in heatlosses-{automatic,initial-geometry}-source-1; heatlosses-initial-geometry-state-map-1.json records the state maps. initial-radius-*, initial-geometry-*, and verify-initial-geometry-*.py retain the analytic reproduction, OMC C, red/green checks, abandoned corrections, control binary hashes, requests, and comparator evidence. All paths in this section are below .git/multibody-campaign/rolling-wheel unless prefixed target/msl. initial-geometry-full-delta-1.json records the completed cohort delta and artifact hashes. No baseline was promoted.

SMPM comparison boundary: no passing credit

SMPM_Braking is recorded as a reference-accuracy limitation under SPEC_0033 §6a. This closes its classification as an actionable compiler refinement counterexample, not its numerical accuracy obligation. Neither solver’s voltage trace has been proved accurate. The full model stays in the next 566-model run and earns no strict-high or certification credit.

The source identity, independent analytic reproduction, unchanged original trace, OMC refinement disagreement, and failed 1e-12 diagnostic below are the basis for this disposition. They reject using the default OMC voltage as an accuracy oracle. In particular, the two refined OMC traces still have five deviating channels; they are not convergence evidence.

The additional native BDF probe stiff-voltage-native-3 measures 47 accepted steps and 2,000 output samples at the same 1e-6 tolerances. After the hard-stop fix, the maximum accepted-point voltage error is 5.68e-7 V. At intermediate points, reconstructing voltage from the approximate current still gives 1.1509 V maximum error, whereas L * interpolate_dy(t) has maximum error 3.00e-7 V. The maximum derivative defect is 3615.62 A/s. This localizes an outstanding numerical issue to amplified observation of the continuous extension; replacing voltage by the polynomial derivative without satisfying the original circuit equations would not be a valid compiler fix.

Gate schema 6 records the reviewed 5→6 boundary, its evidence commit 0de3f29c0ae9440061bef21c7eb9eb4650407015, the original full run multibody-automatic-basis-full, 166 strict-high before/after, and 20→21 excluded models. The exact exclusions-file digest is pinned. The previous 4→5 migration remains part of the checked history. Every existing baseline floor and certified-model roster is unchanged; this is not a baseline promotion. Tests reject missing/forged history and lowered unrelated ratchets.

Validation: 22 xtask baseline tests, 97 quality-gate tests, five promotion metadata tests, Clippy for both affected crates/all targets, and workspace format checks pass. The fixed canary smpm-reference-boundary-canary-1 has nine compared/high models, 175 high initial channels, eleven existing failed models, no skipped/missing traces, and zero phase or band changes against smpm-hard-stop-canary-1. smpm-reference-boundary-canary-delta-1.json records the delta. All prior artifact hashes were verified and retained in smpm-triage-evidence-2.json, with the native derivative probe added (204 artifacts). No compiler or numerical method changes are part of this boundary.

The remaining obligations are a justified error bound for amplified outputs and an independently converged full-model reference. MultiBody regression work resumes with this limitation explicitly recorded. The previous full-run counts below remain historical evidence until the next named complete run.

Automatic-selection full sweep regresses; electrical triage remains open

The complete 566-model run multibody-automatic-basis-full at 5ed8581991cc31418d24f41856f916b5d4c1005d fails the soundness gate. It has 187 raw simulation completions, 167 compared models, 166 strict-high models, one near model, 20 reviewed exclusions, zero missing traces, and zero typed non-identifiable traces. All 22,192 compared initial values are high. Four trajectory channels deviate, all in Modelica.Electrical.Machines.Examples.SynchronousMachines.SMPM_Braking. The near model earns no passing credit. The run uses eleven workers and the ordinary phase budgets; no baseline is promoted.

MultiBody is 20/42 strict-high, down from 25/42 in multibody-single-anchor-full. RevoluteConstraint is restored, but six previously high models are lost:

Five newly high models elsewhere do not discharge these regressions. The per-model delta is automatic-basis-full-delta-1.json under .git/multibody-campaign/rolling-wheel/ (the evidence directory below).

The electrical discrepancy follows an ill-conditioned observable

SMPM_Braking has 584 high channels and four deviating channels, with all 588 initial values high. Three deviations are aliases of smpm.lszero.v; the fourth is voltageQuasiRMSSensor.ToSpacePhasor1.zero. MSL’s SpacePhasor equations, the model’s 1 MOhm grounding resistor, and OMC generated equations 623–635 give, with i0 = smpm.i_0_s:

grounding.i = -3*i0
grounding.v = -3*Rg*i0
vzero = mean(vAC) + 3*Rg*i0 - mean(rs.v)
Lzero*der(i0) = -vzero

The winding resistances are equal, so mean(rs.v) = -Rs*i0. Thus an error in the small zero-sequence current is amplified by 3*Rg + Rs in the observed voltage. Evaluating the displayed voltage identity on the two original traces gives maximum absolute residuals of 5.23e-8 V (Rumoca) and 6.61e-8 V (OMC). Rumoca’s zero-sequence voltage spans approximately [-0.00984, 0.01111] V; the original OMC trace spans [-3.0084, 3.1138] V. This locates a numerical accuracy concern without proving either entire trajectory correct.

The existing xtask-generated OMC executable was inspected directly and reused for bounded diagnostics. Original source, executable, official reference, Rumoca trace, comparison thresholds, and gate tolerances remain unchanged. Only copied diagnostic initialization files request other tolerances:

The five OMC self-comparison deviations are the same four voltage observations and der(smpm.i_0_s). Consequently these two diagnostic traces do not establish convergence of those observables. At 1e-12, OMC reports diode chattering near 0.523893568 s and times out after 60 seconds. That failed attempt is retained; its partial CSV is not a complete reference. At the original tolerance, -noEquidistantTimeGrid still produces voltage excursions exceeding 2 V, rejecting an explanation based only on output interpolation. This flag emits accepted integration steps according to the OMC runtime documentation.

An analytic reduction exposes the same risk in both solvers

The independent, non-switching reproduction StiffVoltage.mo is:

model StiffVoltage
  parameter Real R = 3e6;
  parameter Real L = 0.1/(100*3.141592653589793);
  parameter Real omega = 100*3.141592653589793;
  Real i(start=0, fixed=true);
  Real v;
equation
  L*der(i) = v;
  v = 100*sin(omega*time) - R*i;
end StiffVoltage;

For A=100, w=omega, and D=R^2+(w*L)^2, its exact voltage is A*w*L*(R*cos(w*t)+w*L*sin(w*t)-R*exp(-R*t/L))/D. The steady voltage amplitude is approximately 3.33e-6 V. Over 0–0.02 s with both requested tolerances at 1e-6 and output interval 1e-5 s:

OMC’s accepted-step output still has maximum voltage error 0.4878 V. Its algebraic voltage equation is satisfied to roundoff. The reduction therefore rejects both “OMC is necessarily the accurate oracle” and “Rumoca’s smaller MSL voltage excursions alone prove correctness.” State error control does not by itself bound this amplified observable. The reproduction remains a numerical accuracy frontier, not a passing golden model.

Evidence includes smpm-source-1/ir-{flat,dae}.json, smpm-reference-diagnostic-1.py, smpm-reference-intermediate-1.py, the smpm-omc-refinement-* receipts/comparisons, smpm-omc-output-1, and stiff-voltage-1/analytic-comparison.json. smpm-triage-evidence-1.json binds the source, executable, traces, diagnostics, failed attempt, and full cohort artifacts by SHA-256.

Disposition at the original investigation: no exclusion or tolerance policy was introduced, and no diagnostic earns cohort credit. The next obligation is to establish an accuracy bound for the amplified observable and distinguish nonlinear solve error from integration error using the analytic reproduction. The subsequent comparison-boundary disposition is recorded above. Combined verify quick/full, baseline promotion, push, and PR publication have not run.

Restore the BDF refresh policy when shortening a step

Driving the exact same scalar ODE directly through Diffsol’s sparse BDF and backtracking Newton solver reproduces Rumoca’s dense-output error, without Modelica lowering, algebraic projection, FMI, or the trace sampler. Of 47 accepted steps, the large voltage error is isolated to the shortened final step: -0.4399047825 V at 0.02 s instead of the analytic 3.333333333e-6 V. Other accepted endpoints have maximum voltage error 5.68e-7 V. Intermediate continuous-extension samples can still have much larger error.

Bdf::handle_tstop updates the step coefficient and difference history but omits the ordinary coefficient-dependent Jacobian refresh policy. Newton therefore uses a factorization for the preceding step coefficient. The fix invokes that existing policy after clipping, retaining the existing thresholds, convergence checks, and RHS-Jacobian reuse. No Modelica model or observable names enter production code. SPEC_0038 / ME-INT-004 records the numerical method obligation.

The direct regression shortened_bdf_step_preserves_the_stiff_voltage_solution fails before the fix with the above -0.4399 V result and passes afterward at unchanged 1e-6 integration tolerances. The separate diagnostic now gives 3.245663075e-6 V at the final endpoint, an 8.77e-8 V error, with the same 47 accepted steps. Across all accepted endpoints the maximum voltage error falls from 0.4399 V to 5.68e-7 V. The 2,000 intermediate dense-output samples still reach 1.1509 V error; the hard-stop repair does not establish their accuracy.

Validation passes: 11 Diffsol-wrapper tests, 674 core tests, six spec-budget tests, wrapper Clippy, and workspace formatting. The fixed 20-model canary smpm-hard-stop-canary-1 has the same nine compared/high models, 175 high initial channels, eleven existing failures, zero skips/missing traces, and zero phase or band changes against multibody-automatic-basis-canary-2. The ordinary focused run smpm-hard-stop-focused-1 retains SMPM_Braking’s four deviating channels; its Rumoca trace is byte-identical to the original full-run trace. The partial harness exits zero, but the comparator remains near, so this is explicitly not counterexample closure.

Before/after native probes and metrics are stiff-voltage-native-{1,2}; verification logs use the smpm-hard-stop- prefix. The next numerical frontier is the discrepancy between accurate accepted steps and the amplified observables computed from their continuous extension. The default OMC reference has its own accuracy limitation, so reproducing its voltage noise is not an acceptable target.

Automatic independent-state selection reaches the original trace

The ordinary DAE-to-Solve entry now constructs a source-bound independent state candidate when reduction retains more constrained coordinates than the source differential dimension. Selection uses the shared tensor residual lowerer and directional AD, bounded minimum-norm trial corrections, and a numerically full-rank dependent complement. always coordinates remain independent and never coordinates remain dependent; source preferences order the remaining choices. Trial points never become initial equations, runtime seeds, or OMC-derived overrides. Original initialization and runtime numerical checks remain mandatory. SPEC_0007 / STRUCT-T07 owns the transformation; SPEC_0029 assigns numerical matrix operations to rumoca-eval-solve.

The original RevoluteConstraint DAE completes through this automatic path, without a supplied coordinate set or OMC values. The native comparator checks 918 shared trajectory channels and 918 initial values, all high, with zero minor, deviating, or severe channels. Maximum channel bounded normalized L1 error is 2.74724e-5. The uninstrumented probe takes 16.7448 seconds for build, simulation, and trace serialization together; its 501 samples cover 0–10 seconds with unchanged 1e-6 tolerances and a 12-second solver budget. Evidence is original-automatic-{simulation,comparison}-2 and its source, driver, and trace under .git/multibody-campaign/rolling-wheel/.

That total is not a 14-second simulation-watchdog measurement. The normal worker gives Solve, backend preparation, initialization, and Sim separate phase budgets. An earlier progress update conflated these; the ordinary-worker result below is the authoritative phase-budget evidence. The perf diagnostic of the same executable completes in 16.6391 seconds. Leading self costs include structural dependency projection (5.42%), block output-index construction (3.50%), and assignment-queue operations (3.28%). Profiling records are in original-automatic-perf-2; these percentages alone do not locate a timeout.

Connecting the automatic path exposed general defects, reproduced in existing small core fixtures and repaired at their owners:

• Causal-definition admission required identical types, so Real normal[3] = {0,0,1} was not seeded. Writing Real literals alone made the contact fixture pass, isolating the first divergence. The unchanged Integer-literal fixture now passes using the shared assignment-compatibility query and typed Integer-to-Real coercion, as required by MLS §10.6.13 / ARR-009.
• Blanket preservation of every start attribute retained inherited zeros for algebraic observations, leaving atan2 AD at the undefined origin. A DefaultStart provenance hypothesis also failed because inherited defaults may carry source provenance. The final trial policy retains state, fixed, always/prefer, and explicit override guesses, and evaluates required acyclic definitions for the other algebraics. Original constraints still determine actual initialization.
• Candidate aggregate states lost nonzero source starts. They now project original start expressions, including parameter dependencies, scalar broadcasts, and matrix component order, with fixed=false. Formal derivative tensors retain their unfixed zero guesses (MLS §8.6).
• Required algebraic states had no formal successor when only an alias was differentiated. Tensor-offset refinement now supplies that successor while independently preserving every signature inequality, matching equality, and differential dimension (MLS §4.8.7.1).
• Reconstruction copied a model-level LinearSolve outside its aggregate function owner. Replay now uses the existing shared pure-function owner; function-scoped solves retain their typed operation. Existing non-singular contact tests pass, while the singular auxiliary fixture still refuses.
• Fixed algebraic initialization rows previously claimed every unknown. Planning now follows each exact matched definition and start-expression dependency and prioritizes lower-degree rows. Complete residual and numerical rank checks remain required; structural degree is not a regularity proof.

The bilinear tensor fixture requested fifteen always coordinates in a three-dimensional system. Its original source remains an explicit rejection test; legal positive variants require the three independent coordinates and prefer the dependent tensor observations. OMC accepts the original but selects only three matrix elements, recorded in automatic-always-omc-1, rather than retaining all fifteen requirements. That acceptance does not satisfy MLS §4.8.7.1. OpenModelica ticket 3689 also records ignored always requirements as a compiler defect. No failing model is removed from the corpus.

The first ordinary-worker run, multibody-automatic-basis-revolute, also passes the trace comparator: 918 high trajectory and initial channels, no skips, missing traces, exclusions, or nonidentifiable traces. Solve takes 8.4704 s, backend preparation 3.3352 s, initialization 0.0552 s, and simulation 1.0752 s. The fixed twenty-model multibody-automatic-basis-canary preserves all prior phase and agreement bands: nine compared high, 175 high initial channels, eleven existing failures, and zero skipped/missing/excluded/nonidentifiable traces. Its receipt binds worktree digest 46ba7f678c8bb55da0b1ee169e81dad920c3e9ac270e4c940c7735e1e61752fb.

Inspecting the successful worker’s state metadata exposed an additional handoff defect: the four-coordinate candidate reached the runtime with six states. Its consuming preparation reran source state promotion, adding joint.phi and joint.w alongside the aggregate that already represented them. A one-state always fixture reproduces the same unwanted reconstruction. The candidate now retains its constructor-issued structural analysis and moves the exact checked root into preparation, transferring original initial-value obligations without repeating state selection. The regression proves unchanged declarations, one state, retained fixed-value pins, and structural ownership. All 674 core tests pass after this fix. The next worker receipt measures its four-state execution; the earlier six-state result is retained as evidence, not silently relabeled.

Architecture checks also exposed new runtime assertions and an unchanged dependency pin. Formal stage coordinates now carry their checked declaration views into Solve; fixed-value incidence matches typed identities directly. Numeric trial data and typed-program construction propagate explicit failures instead of panicking on unavailable inputs. The totality-debt ceiling is unchanged. The exact dependency inventory adds the specified external nalgebra numerical library and still forbids every upward Rumoca dependency.

Final ordinary-worker evidence is target/msl/multibody-automatic-basis-revolute-2: the trace records exactly four runtime states, and the comparator again checks 918 trajectory channels and 918 initial values, all high. Maximum channel bounded normalized L1 is 6.07710e-6. Solve takes 8.2368 s, backend preparation 3.3438 s, initialization 0.0559 s, and simulation 1.0344 s. Every phase fits its unchanged budget. The state-name-set comparator is unmeasured for this aggregate representation; the four-state count comes directly from the trace, not from that comparator’s zero-count summary.

multibody-automatic-basis-canary-2 again preserves every phase and band from the same twenty targets: nine high comparisons, 175 high initial channels, eleven existing failures, zero skipped/missing/excluded/nonidentifiable traces. The delta binds worktree digest 2737630234c0e37186a8bf343130bb4e90a5838f019be0e2617777c16cd38490. Validation passes 674 core tests; 194 evaluator, 190 DAE, 144 Solve, 217 structural, and 109 GALEC library tests; 243 architecture tests; six spec gates; affected-library and core Clippy; formatting and whitespace checks. Evidence is bound by rolling-wheel/automatic-basis-evidence-1.json.

This restores the original model in focused validation; it is not a new full-cohort coverage claim. The next named-commit Tier 2 sweep must preserve all previously high models before breadth resumes. Release quick/full, baseline promotion, push, and a PR remain pending.

RevoluteConstraint: original equations and a supplied independent basis

The original formal system now has numerical evidence against OMC, beyond the reduced rotation fixture. OMC’s generated initialization XML supplies exact alias records, including initAngle.angle[2] = derd[2].u; those records map its CSV values to source coordinates in a diagnostic. Rumoca’s source-bound kernels reconstruct eliminated intermediate variables while holding every mapped OMC value fixed. At times 0 and 7.64, all 2,640 formal residuals are below 3.9e-13. No model-name or OMC-dependent selection logic was added to production.

Numerical selection on the typed AD matrices, with the source’s StateSelect.always requirements enforced, chooses the same four coordinates as OMC: joint.phi, joint.w, initAngle.angle[2], and its first derivative. The dependent stage matrices have full numerical rank at both sampled points. The highest-stage condition estimate reaches 8.02e7 at 7.64; lower-stage estimates stay below 4.9e4. These are sampled diagnostics, not a global regularity certificate. Evidence: rolling-wheel/original-basis-selection-1.json.

The first end-to-end experiment failed before integration: generated formal derivatives had no start attribute, and strict runtime preparation refused $formal_derivative.1.joint.frame_a.r_0. The proposed aggregate state also lacked a seed. The constructors now use the existing compact shaped_zero helper and explicitly retain fixed=false. This follows SPEC_0007 / STRUCT-T07 and MLS §8.6: unfixed starts are iteration guesses. Source attributes and initial equations remain authoritative. Two strengthened regressions fail with MissingValue before the fix and verify numerical seeds, tensor shape, and unchanged source initialization after it.

A temporary test-only export let the existing simulator execute the original model’s constructed four-state candidate. The export probe was removed; no production admission API was opened. The simulation uses original source initialization and fresh zero guesses for generated variables, with no OMC value overrides. BDF completes 501 samples over 0–10 seconds with unchanged 1e-6 relative/absolute tolerances and a 12-second solver budget; total preparation plus simulation time was 11.45 seconds.

The repository trace comparator reports high agreement for all 918 shared channels, with zero minor, deviating, or severe channels. All 918 initial-condition checks are high. Maximum channel bounded normalized L1 error is 2.75e-5. This is one explicitly supplied-basis experiment, not a normal-path MSL pass or a cohort metric. Evidence: original-candidate-simulation-2.jsonl, original-candidate-comparison-1.json, and the hashed traces and export probe in formal-guesses-evidence-1.json.

During numerical reconstruction, two angle-sensor Jacobian rows were non-finite at the deliberately zero matrix guess. The diagnostic first solved the other constraints; the final complete Jacobian is finite, full rank, and all residuals and assertions pass. Those intermediate rows were not omitted from final validation or from the simulation. This is a starting-guess issue, not evidence that a singular or non-finite basis may be admitted.

Validation passes: all 29 focused formal-construction tests, 669 compiler-core tests, 217 structural tests, 243 architecture checks, six specification gates, structural/core Clippy, formatting, and whitespace checks. The fixed target/msl/multibody-formal-guesses-canary has no phase or band delta from multibody-formal-kernels-canary: nine compared models and all 175 initial condition channels remain high; eleven existing failures remain visible. Skipped, missing, excluded, and nonidentifiable counts are zero. The receipt formal-guesses-canary-delta-1.json binds base commit 1baf8e68e581d8e4d56c26f6cfed50b64dd23abe and working-tree digest 64bbb7d1fbf727fea3879286a42a2e5d36478e09080513e9c5e5e8aebe91d6a0.

The remaining production work is automatic numerical basis selection, admission, initialization/reconstruction ownership, and FMI state mapping. The normal-path RevoluteConstraint EX002 regression remains open. No new Tier 2 coverage claim, release quick/full run, baseline promotion, push, or PR is made here.

RevoluteConstraint: shared tensor kernels for stage Jacobians

lower_formal_derivative_stages now lowers source-bound formal stages through Rumoca’s shared typed expression lowerer and its directional AD. The borrowed analysis product retains complete tensor captures, equation owners and body ordering, compact domains, source provenance, and reached function assertion predicates. Original initialization and assertions remain attached through the formal root. It grants no numerical regularity, prepared-model, integration, or FMI authority. SPEC_0007 / STRUCT-T07 and SPEC_0029 own this boundary.

The original RevoluteConstraint now constructs kernels for all 2,640 formal residuals: stages -2/-1/0 contain 176/256/2,208 rows in 36/68/663 equation owners. Every owner has a checked directional program. This is construction evidence, not a successful simulation or numerical rank certificate.

This exposed three general defects in the shared typed lowerer:

• Frames.relativeRotation passes Real[3] under a different declaration-owned type ID from its parameter. DAE already proves complete value-type equality; Solve incorrectly rejected unequal IDs. The fix preserves the target identity while accepting equal complete value types. A checked-DAE regression fails before the fix and verifies scalar/vector/matrix values and AD afterward.
• Visualizers.Internal.Lines repeats orientation matrices for component arrays. The old binder-independent comprehension path fed a tensor to scalar-only fill. A compact map now repeats the whole tensor. Comprehensions also use their own declared domain when a nested body reads only an enclosing binder. Checked-DAE tests verify both axes and AD; source fixtures that simplified away the failing construct were not accepted as regressions.
• vector([lines[i,1,:];0]) already has a checked DAE owner, but typed Solve lowering was missing. It now uses compact singleton-axis projection, or a one-element aggregate for scalar input, preserving element order and AD.

These fixes follow MLS type-compatible expressions (§6.7), dimensionality conversion (§10.3.2), and array constructors (§10.4.1). No model-name branches or tolerance changes were introduced. Nested-call and model-coordinate capture logic is shared with existing event lowering.

For the reduced rotation model, the new evaluator’s 129 residuals and full 129-by-131 AD Jacobian agree with an independent complex-step evaluation of the canonical equations at cosine 1, 1e-3, 8.516739322852923e-6, 1e-7, and 1e-9. Maximum residual and Jacobian differences are below 9e-16. This includes the saved failure neighborhood. A diagnostic selector using the actual typed AD matrices chooses q[2], w[2] at all five points; this does not certify a trajectory or implement production state selection. Evidence: rolling-wheel/formal-kernels-comparison-2.json and the final probe artifacts listed in formal-kernels-evidence-1.json.

Validation passes: all 144 Solve unit tests, 669 compiler-core tests, 243 architecture checks, six specification gates, Solve/core Clippy, and workspace formatting. The new regressions cover tensor and loop residual values and AD, source coordinates and explicit time, call assertion preservation and explicit refusal of unsupported map-scoped assertions, equal complete value types, nested comprehension domains, and vector conversion. Logs use rolling-wheel/formal-kernels-*.

The fixed target/msl/multibody-formal-kernels-canary has no phase or band delta from multibody-formal-stages-canary: nine compared models and all 175 initial condition channels remain high; eleven existing failures remain visible. Skipped, missing, excluded, and nonidentifiable counts are zero. The receipt formal-kernels-canary-delta-1.json binds base commit 94dee5f9c8af43abff9789e129492bbd7d3ade4e and working-tree digest 8cd650ce731ecec4fcb17641f7a3cb7ef9c8191e01df6b5a833e8a881f849f79.

The next obligation is numerical selection on these stage kernels, followed by initialization, reconstruction, and runtime/FMI state mapping. The original RevoluteConstraint EX002 regression remains open. No full-cohort claim, release quick/full, Tier 2 sweep, baseline promotion, push, or PR is made here.

RevoluteConstraint: derivative stages for numerical state selection

FormalDerivativeView::stages now exposes source-bound equation and coordinate owners grouped by derivative order minus certified tensor offset. Construction retains a compact equation-owner map alongside the existing tensor-coordinate map. These borrowed views partition the formal system, retain domains and provenance, and expose structural freedoms without adding scalar owners or authorizing numerical execution. The contract is SPEC_0007 / STRUCT-T07.

The production views give these equation/coordinate counts, from lowest to highest derivative stage:

A numerical diagnostic consumes the actual issued reduced stages and eliminates noncandidate columns before pivoted QR selects among original source states. It automatically chooses q[2], w[2] at five tested points from cosine 1 down to 1e-9. At the saved failure’s cosine, the stage condition estimates are at most 1.75e6, whereas the earlier complete reconstruction matrix was around 1e17. This explains why its dense condition estimate alone is an unsuitable admission test. At angle -pi/2, a negative control detects rank loss in every stage and admits no basis. These are local numerical diagnostics, not runtime state selection or new trace results. Evidence: rolling-wheel/staged-basis-2.json and staged-rank-boundary-1.json.

A fresh OMC replay of the original model covers 502 rows over 0–10 seconds. Its selection matrices retain freeMotionScalarInit.derd[2].u and freeMotionScalarInit.angle_d_2 throughout, alongside the two fixed joint states. Rumoca constructs that corresponding angle/angle-derivative proposal with all 2,648 equations matched, all 36 source functions, and all four fixed declarations preserved. Evidence: omc-revolute-state-selection-1.{log,json} and formal-stages-omc-basis-1.jsonl. Numerical admission, initialization, runtime, and FMI mapping are still required; the original Rumoca regression remains open.

All 662 compiler-core tests, 217 structural tests, 243 architecture checks, six specification gates, structural/core Clippy, and formatting pass. The five new tests check exact partition coverage, source ownership, no later-stage dependencies, loop and matrix domains, empty tensors, explicit time and function dependencies, and unconstrained or zero-dimensional systems.

The fixed target/msl/multibody-formal-stages-canary has no phase or band delta from multibody-state-candidate-canary: nine compared models and all 175 initial condition channels remain high; eleven existing failures remain visible. Skipped, missing, excluded, and nonidentifiable counts are zero. The receipt formal-stages-canary-delta-1.json binds base commit 69410b809acd65126e1ef7382e869d8ee1a8c2c0 and working-tree digest ee2c4169e3f27c3b5ef6998ef902c0a68916c5ad15ee82031445bfeba1aba831. No new cohort coverage is claimed. Release quick/full and Tier 2 were not rerun while execution remains unfinished; there is no promotion, push, or PR.

RevoluteConstraint: checked state-coordinate candidates

The formal derivative system now constructs an inspectable state-coordinate candidate. Each selected source scalar is bound to its formal derivative successor. Construction checks the formal dimension, distinctness, tensor bounds, required and forbidden source state preferences, and complete structural matching. One unfixed aggregate state and two tensor projection equations represent a nonempty selection; zero-dimensional systems add neither states nor initial conditions. Original tensors, functions, assertions, attributes, and initialization obligations survive the same checked root replay. SPEC_0007 / STRUCT-T07 was amended before implementation.

The reduced rotation source accepts both the supplied q[2], w[2] basis and a q[2], der(q[2]) proposal, with 133 matched equations. A source-artifact probe also constructs the original RevoluteConstraint’s four-coordinate proposal: joint.phi, joint.w, freeMotionScalarInit.initAngle.angle[2], and bodyOfConstraint.body.w_a[2]. All 2,648 equations match; all 36 source functions and four fixed source declarations remain. The original source is unchanged. Names select coordinates only in this diagnostic, not in production code. Evidence is rolling-wheel/state-candidate-source-1.jsonl and its retained probe.

This product cannot be passed to numerical execution. Structural matching does not establish a regular basis, and automatic selection, initialization admission, basis changes, and FMI mapping remain pending. The original RevoluteConstraint simulation failure is still open; the earlier normal-budget replay below remains the latest execution evidence. This construction adds no MSL coverage result.

Validation passes: eight new focused tests, all 657 compiler-core tests, 217 structural tests, 243 architecture checks, six specification gates, structural and core Clippy, and workspace formatting. Tests check the actual matrix value and derivative projection expressions, source preservation, invalid selections, hard state preferences, and a structurally singular angle-only proposal. Logs use rolling-wheel/state-candidate-*.

The fixed target/msl/multibody-state-candidate-canary has no phase or band delta from multibody-tensor-refresh-canary: nine compared models and all 175 initial condition channels remain high; eleven existing failures remain visible. Skipped, missing, excluded, and nonidentifiable counts are zero. The receipt state-candidate-canary-delta-1.json binds base commit 80ed1fdd42b1a074f1e50b2258d4be2f068a8476 and working-tree digest 51af999509e577769190a88ee744234a3366c97f990593d09437d08543b96ca1. Release quick/full and a new Tier 2 sweep were not run while independent-state execution remains unfinished. No promotion, push, or PR is claimed.

RevoluteConstraint: coupled execution and exact refresh dependencies

A diagnostic source exporter now preserves the formal construction’s 23 tensor and scalar equation owners and both initial equations, adding the explicit two-coordinate basis q[2], w[2] used by the earlier handwritten prototype. Normal compilation and runtime execution of these automatically differentiated equations pass eight value checks and sixteen state JVP checks, including guesses with every algebraic coordinate perturbed by 1e-4. The four tested cosine values range from 1e-3 to 1e-9. After observation refresh, all 131 formal coordinates agree with the analytical values within 8.881784197001252e-16. Both BDF and RK produce 101 samples whose 47 source channels agree with the saved OMC trace within 2.6423307986078726e-13. Evidence is rolling-wheel/coupled-generated-comparison-3.json and its retained probes. This is a diagnostic with a supplied basis, not automatic state selection or a new MSL coverage result. OMC’s generated _04set.c also retains two states, but selects them dynamically from separate angle and rate candidate sets.

The comparison exposed a general refresh-planning defect. An exact assignment to one tensor component traversed every read in its shared residual program, pulling unrelated sibling equations into the derivative kernel. The minimal case der(x)=a[2]; a={y,-x,0}; y*y*y=x therefore solved for y on every RHS request even though the required assignment is just a[2]=-x.

collect_dependency_closure now follows the same checked assignment shape as an exact assignment already admitted by the full plan. Coupled or uncertified projection stages retain their existing dependencies. The source tensors, equations, algebraic observation plan, and tolerances remain authoritative. This implements SPEC_0007 / SOLVE-C56 using the shared dependency query owned by SPEC_0029 / SPEC_0041. The focused regression proves the unnecessary solve is deferred, state values and AD remain correct, requested observations are solved, and genuine value/coefficient/coupled dependencies remain required.

For the generated rotation diagnostic, the derivative plan shrinks from 103 to 69 projection blocks and omits the 13-variable second-derivative coupled solve. The complete observation plan retains all 103 blocks, including that solve. The final replay retains the value, JVP, and source-channel accuracy above. The earlier dense full-Jacobian condition estimate therefore does not by itself predict failure of this structured reconstruction on the tested state directions.

Validation passes: 649 compiler-core tests, 191 evaluator tests, 503 solver tests, 243 architecture checks, evaluator/core Clippy, formatting, and whitespace. Logs use rolling-wheel/tensor-refresh-*; the first focused run reproduces the bug. The fixed target/msl/multibody-tensor-refresh-canary has no phase or band delta from multibody-formal-construction-canary: nine compared models remain high, all 175 initial-condition channels remain high, and eleven existing failures remain visible. Skipped, missing, excluded, and nonidentifiable counts are zero. tensor-refresh-canary-delta-1.json binds base commit 319d0d99e067e8d9f01cca87f6ded40fc52f0fa7 and working-tree digest d009613b065f8b26e7cb97e8552557c38870316fad8deb9baf2d612c87c04b08.

The fresh normal-budget original-model run in target/msl/multibody-tensor-refresh-revolute still fails with EX002 on bodyOfConstraint.body.z_a[3]: sensitivity residual -1.164153e-10 against the unchanged 1e-10 tolerance. It produces no comparable trace, so its gate correctly reports parity unmeasured. Automatic independent-state construction and that regression remain open. No new Tier 2 cohort result, release quick/full run, baseline promotion, push, or PR is claimed.

RevoluteConstraint: construct coupled formal derivative equations

construct_formal_derivatives now builds an inspectable system from the original source and certified tensor-uniform offsets. One checked root replay retains source values, attributes, initialization, assertions, functions, and compact equation domains. It reserves complete derivative tensors and appends the required differentiated equations, using the existing tensor differentiation rules with a separate formal-coordinate map. Source-solution zero anchors and causal algebraic definitions cannot erase or replace those derivative unknowns. No auxiliary inverse is introduced unless a linear solve already occurs in the source. The governing contract is SPEC_0007 / STRUCT-T07.

The final construction replay over the original source artifacts produces:

These equations are now emitted, unlike the offset-only counts below. Function counts remain 0 and 36 respectively, equal to the source counts. Both reduced initial equations remain; the original model’s declaration-based initial attributes also survive. Evidence: rolling-wheel/formal-construction-source-3.{jsonl,log} and its retained probe.

An independent NumPy evaluation of the emitted reduced expression graph checks all 129 equations at four cosine values down to 1e-9. The maximum residual is 8.881784197001252e-16. Perturbing each of the 131 coordinates changes the residuals. All 47 source values reconstructed at the 102 saved OMC trace rows agree within 4.996003610813204e-16, and the generated equations hold there. formal-equations-comparison-2.json binds this diagnostic. This evaluates a known solution and its formal derivatives; it is not a new integration or JVP result.

A separate complex-step Jacobian and pivoted-QR diagnostic exposes the next obligation. The full unscaled reconstruction Jacobian is severely ill-conditioned near the Euler-angle singularity: at cosine 1e-3, both the QR-selected basis and the diagnostic q[2], w[2] basis have condition numbers around 8.35e11. At smaller cosines the smallest singular direction is below reliable double precision resolution. The resolved near-singular direction at 1e-3 is dominated by second angle derivatives, first rate derivatives, and associated rotation derivatives, with smaller visible-rate components. Changing a pair of states alone does not settle this issue. Evidence is formal-basis-1.json and formal-conditioning-2.json. Dependency closure, structured coupled solving, and accuracy of the required outputs need investigation before execution.

All 647 compiler-core tests, 217 structural tests, 243 architecture checks, and six specification gates pass. The six new construction tests cover source and tensor preservation, pinned-value derivative obligations, compact loop binders, function substitution, fixed starts, state preferences, assertions, and typed refusal of incompatible orders, nonsmooth differentiation, and unsupported reinitialization mappings. Structural and core Clippy, workspace formatting, and whitespace checks pass. Logs use rolling-wheel/formal-construction-*.

The fixed target/msl/multibody-formal-construction-canary has no phase, simulation, or band delta from multibody-tensor-offsets-canary. Nine models remain high, all 175 initial-condition channels remain high, and eleven existing failures remain visible. There are no skipped, missing, excluded, or nonidentifiable traces. formal-construction-canary-delta-1.json binds HEAD 2e2adf2f161eeb24a73571f8920b1802638387f4 and working-tree digest f38cf07fafecbe85bcf8be017977ded6b25b3c022b1ec6148806325ef24e3dc4.

The construction product is not a prepared numerical DAE and is not selected by the simulation pipeline. Independent-state execution and the original RevoluteConstraint regression remain unresolved. Release quick/full and a new Tier 2 sweep were not run while these obligations remain open; no new cohort pass, promotion, push, or PR is claimed.

RevoluteConstraint: differential orders for complete source tensors

The source differential analysis now checks whether its scalar derivative orders can be raised to one order per complete variable declaration and continuous equation owner. The bounded closure preserves every signature inequality, matched equality, and formal dimension; an independent certificate checks the result. Canonical tensors and source equations remain unchanged. An incompatible uniform refinement returns no tensor result while retaining the valid scalar analysis, rather than declaring the source singular. The governing contract is SPEC_0007 / STRUCT-T07.

The production analysis replay yields these planned differentiated-system counts:

These are counts implied by certified orders, not equations emitted or executed by a new compiler transformation. The first two dimensions agree with OMC’s state inventories. Numerical regularity, coupled construction, initialization, and independent-state selection remain outstanding; the original RevoluteConstraint regression is still open. The transformed input’s separate retained manifold remains as described below. Evidence is rolling-wheel/tensor-offsets-source-2.jsonl, its retained probe and driver, and the independent diagnostic tensor-offsets-uniform-1.json.

All 641 compiler-core tests and 217 structural tests pass, including exhaustive two-by-two signature/owner-partition comparison against an enumeration oracle. The final 11 differential-analysis tests pass after the oracle cleanup. Controls cover empty owners, incompatible component orders, unchanged source tensors, invalid partitions, and checked overflow. The 243 architecture checks, six spec gates, structural all-target/all-feature Clippy, compiler-core Clippy, and workspace formatting pass. Logs use rolling-wheel/tensor-offsets-*.

The fixed target/msl/multibody-tensor-offsets-canary has no phase, simulation, or band changes from multibody-init-algebraic-canary: all nine compared models remain high, all 175 initial-condition channels are high, and there are no skipped, missing, excluded, or nonidentifiable traces. Eleven existing failures remain visible. The receipt rolling-wheel/tensor-offsets-canary-delta-1.json binds HEAD 50382b187a4827c371932f9d4e32341eaed9ce2f and working-tree digest db1d2d5c9cda34f6d6d49408c6a8d54919f7d454321945cdc46ebe05522297db.

This analysis is not yet used by the numerical kernel and adds no cohort pass. Release quick/full and a new Tier 2 sweep were not run at this checkpoint: automatic coupled-state construction is unfinished. No baseline promotion, push, or PR is recorded here.

RevoluteConstraint prototype: preserve source initialization through algebraics

The manually constructed two-state coupled rotation equations passed all eight value evaluations and sixteen JVP directions, including perturbed algebraic guesses at cos(angle) = 1e-9, but both integrators refused the original initial equations q[2]=0; w[2]=-2.4. The first divergence was Solve initialization: those algebraic reads were classified as unowned, leaving no projection blocks. The small InitArrayAlias.mo fixture reproduces it without rotation equations. OMC initializes that source to z={0,-2.4}. A diagnostic control stating the equivalent conditions directly on z also initializes in Rumoca; changing the source initial conditions is not the fix.

Solve now indexes both algebraic and derivative matched continuous row views in one ContinuousRowIndex. Initialization follows those source rows transitively with exact tensor components and structured domain points. Visited coordinates close coupled algebraic loops, exposing their external state/parameter reads. Admitted rows use the existing reconstructed-residual total derivative; missing or unsupported ownership still retains reconstruction before residual checking. Fixed states and all surplus equations remain binding. The governing contract is MLS §8.6 and SPEC_0007’s initialization paragraph. No source declarations, equations, or initialization values are rewritten by this change.

The prototype now completes 0–1 s on both BDF and RK. Across 101 samples and all 47 common source channels, its maximum absolute difference from the original reduced model’s OMC trajectory is 2.6423307986078726e-13. Evidence: rolling-wheel/coupled-rotation-{4,5}.log, coupled-rotation-omc-comparison-1.json, and the retained probes/drivers. This is a manually assembled representation test, not automatic independent-state construction or a new MultiBody cohort pass. The initialization runtime’s existing finite-difference total sensitivity is unchanged; replacing that implementation with shared AD remains separate work.

The small alias, coupled array loop, indexed equation family, and parameter-binding controls agree with OMC’s initial values. Their sources, generated code, traces, and analytical error measurements are retained in init-array-alias-omc-1 and init-algebraic-controls-omc-1. Regression tests cover both integrators, conflicting fixed states, and an unsupported discrete dependency whose stale zero algebraic seed must not certify initialization.

Validation: 638 compiler-core tests, 17 final tensor-initialization tests, 46 simulation-lowering tests, 141 Solve-phase tests, and 243 architecture checks pass. Solve all-target/all-feature Clippy, compiler-core Clippy, all six spec gates, workspace formatting, and whitespace checks pass. Logs use init-algebraic-*.

The fixed target/msl/multibody-init-algebraic-canary has no phase, simulation, or band delta from multibody-differential-structure-canary. All nine compared models remain high; there are zero skipped, missing, excluded, or nonidentifiable traces, and all 175 initial-condition channels are high. The other eleven members retain their existing failures. The receipt is rolling-wheel/init-algebraic-canary-delta-1.json; the run binds HEAD 17be389de883177f214b2a741eefe963fd0882f8 and working-tree digest 6f0c4970dc07677d24404bd8f3a742e64db031dcb0d6710956101495c155cb47. This Tier 1 check is not a new cohort coverage number.

The original RevoluteConstraint regression remains open. Automatic construction of the coupled differential system, original-model replay, and the complete cohort preserving prior high models are still required. No release quick/full, new Tier 2 claim, baseline promotion, push, or PR is recorded here.

RevoluteConstraint: checked source differential structure

The structural phase now derives a source-bound differential signature and certifies its maximum-weight assignment and least nonnegative equation/variable offsets. It shares the existing tensor/function scalar projection and canonical continuous-owner ordering. Source declarations, equation families, and initial conditions remain unchanged. The implemented contract is SPEC_0007 / STRUCT-T07; SPEC_0053 still proposes the unimplemented independent-state execution system.

The final public-API replay over the retained source artifacts establishes:

The first two dimensions agree with OMC’s generated state inventories. The last row does not mean the four constraints were discarded: their lower-order forms are retained separately in the prepared system’s manifold. It identifies the redundant differential system that independent state construction must replace. The reduced source signature requires differentiating the rotation constraints twice and the rate definitions once. No numerical inverse or simulation outcome supplies those orders. Evidence is rolling-wheel/differential-structure-real-2.{jsonl,log} and its retained public probe/driver; OMC inventories remain in rate-cancellation-omc-1 and revolute-initial-omc-1/omc_sim_work.

All 213 structural tests pass, including exhaustive comparison of all 19,683 three-by-three signatures against an independent enumeration oracle. Negative certificate, overflow, singular-graph, and hidden pendulum-constraint controls pass. All 633 compiler-core tests pass; the eight source-analysis tests also pass after the final helper extraction. These cover unchanged tensor ownership, loop domains, function permutations, inputs/outputs, tunable versus invariant zeros, and the distinction between structural and numerical regularity. Structural all-target/all-feature Clippy, compiler-core Clippy, and all six specification gates pass. Logs use the differential-structure- prefix under rolling-wheel.

The fixed target/msl/multibody-differential-structure-canary has no phase, simulation, or band changes from multibody-invariant-value-canary: all nine compared models remain high, all 175 trajectory/initialization channels are high, and there are no skipped, missing, excluded, or nonidentifiable traces. The other eleven canary members retain their existing failures. The run records HEAD 928b1e30e306db21e2e98267cc0e5ba6067a31ec and working-tree digest 305703ae6953295248506ecbc2b8221c98e1da3020b25bc52641d6af86c207bd; differential-structure-canary-delta-1.json binds the comparison artifacts.

This is analysis groundwork, not a RevoluteConstraint fix or new cohort result. The analysis is not yet used to construct the numerical kernel. Differentiability, numerical regularity, coupled reconstruction, initialization, and FMI coordinate mapping remain required before it can authorize independent integration states. Combined quick/full and a new complete Tier 2 sweep were not run at this checkpoint: executable state selection remains unfinished and the known Revolute regression remains open. No baseline promotion or PR is claimed.

The architecture scan exposed existing branch cleanup as well as an unexercised matching accessor. The source permutation regression now checks the returned matching coordinate. The retained diagnostic API ReductionSnapshot::manifold_expression_ordinals has concrete out-of-corpus consumers: inspect-revolute-initial-1.rs, inspect-revolute-candidates-1.rs, and inspect-line-force-reduction-3.rs under rolling-wheel. It remains an explicitly documented public diagnostic contract so these inspections can bind retained constraints to the same stalled DAE, without mutating reduction. Coordinate-definition tests move into their own module to satisfy the existing file-size gate. Wire call replay now receives already-destructured call fields from its exhaustive enum match, eliminating a repeated-match unreachable! without changing wire validation or format. All 190 DAE tests, 213 structural tests, eight source-analysis regressions, and 243 architecture checks pass. The DAE totality count is back to its existing ceiling of 60; no ceiling or file-size exemption changed. Final all-target/all-feature Clippy for both libraries, workspace formatting, and whitespace checks pass. Cleanup logs use differential-structure-cleanup- and differential-structure-final- prefixes. The canary above precedes this behavior-preserving wire/test cleanup; it is not a new full-cohort measurement.

RevoluteConstraint: independent coordinates and reconstruction limits

The value/Jacobian callback mismatch hypothesis is disproved. Both production FMI callbacks set the same time and continuous states, then evaluate the same unprojected state vector with algebraic refresh. Manifold correction belongs to the accepted-point path. There is no missing derivative of a state projection inside these callbacks to repair.

The earlier coupled-rate experiment also fails as a complete repair. Keeping rate as three additional states removes the original refusal at the captured angle but still gives -2.4375 instead of -2.4010301622267085 at cosine 1e-7, and zero at 1e-9. Its diagnostic source and output are rolling-wheel/implicit-rate-cancellation-probe-4.{rs,log}. No production change from that experiment remains.

OMC now has an independent build and simulation of the exact tracked RateCancellation.mo fixture. Its generated _04set.c selects one coordinate from each of two three-candidate sets: angular rates and angles. It integrates two independent states where Rumoca retains six. The OMC run completes the 0–1 second interval in 29 steps, with no convergence or error-test failures. All 47 source channels across 102 output rows agree with the analytical motion q={0,-2.4*time,0}, w=rate={0,-2.4,0}, ax=az=0, and the corresponding rotation matrices; the largest absolute error is 7.993606e-15. This is reduced fixture evidence, not a cohort or speed claim. Build, generated C, CSV, and the digest-bound analytical receipt are in rolling-wheel/rate-cancellation-omc-1.

Public evaluator probes then extract the actual retained-manifold Jacobian. The reduced fixture has four independent constraint rows over six coordinates; the saved real model has four over eight. A diagnostic enumeration of square minors identifies these local coordinate sets without using model names:

Solving the constraint tangent equations makes all twelve directions pass the unchanged runtime JVP check: eight reduced-fixture directions and four real-model directions. The reduced results exactly match der(q[2])=w[2] and der(w[2])=0. These are admissible constrained directions, not replacements for the original failing ambient partials. The enumeration is diagnostic only; it is not a proposed production search algorithm. Evidence is constrained-tangent-{matrices,basis,directions,results}-1 under rolling-wheel.

The stronger reconstruction experiment starts dependent coordinates 1e-4 away from the saved point and solves with independent coordinates fixed. At the existing 1e-10 tolerance, the real model’s four reduced AD columns agree with centered differences at three step sizes; the largest absolute difference is 5.655291e-9. However, at cosine 1e-9, the reduced fixture can retain a velocity-constraint residual of 6.295741e-11 and then refuse its q[2] tangent because the dependent acceleration JVP residual is -9.313226e-10. A stricter experimental 1e-12 projection previously stopped on the real model’s floating-point residual floor. Both outcomes remain in the logs; no production tolerance was changed. chart-reconstruction-5.log records every direction, including the refusal, and is not a passing regression test.

Consequently a runtime tangent-projection wrapper is insufficient. The next implementation must construct independent differential coordinates and their reconstruction/derivative equations together, upstream of the expanded ambient RHS. SPEC_0053 records that proposed compiler/FMI contract, including tensor ownership, initialization, basis changes, and the stronger reconstruction tests. It is explicitly unimplemented. Compiler behavior remains c488115f, the Revolute regression remains open, and no new full-cohort result or PR is claimed.

All six specification inventory, citation, and size gates pass in constrained-state-spec-gates-1.log. This checkpoint changes documentation and the proposed contract only; combined quick/full and a new MSL sweep were not run because executable compiler behavior is unchanged.

RevoluteConstraint: cancellation reproduced independently of integration

The EX002 regression remains open at compiler commit c488115f. The saved point at t=7.639844286351321 has its middle initialization angle near -pi/2. A unit angular-velocity direction produces angle-rate sensitivities around 117,416. Recording the actual directional program exposes intermediate values around 66 billion. One contributing sum is -563837.8893896533 + 563837.8893896535; its surviving roundoff reaches the reported residual of -1.164153216573989e-10. The original LU residual is small; re-evaluating the expanded directional expression introduces the disagreement. Increasing an acceptance tolerance would conceal this evidence.

Two attempted repairs were rejected and removed:

• Bounded sensitivity refinement passed its focused tests, including rejection of deliberately inconsistent JVPs, but still refused the real model. It is archived in rolling-wheel/abandoned-sensitivity-refinement-1.
• Preferring source observation lifts before second differentiation restored the rotation-matrix state. Its ordinary worker completed the whole interval with all 918 shared trajectory and initialization channels high, and 1,072 focused tests passed. Replaying the same saved point, however, still failed one direction with the original EX002 residual. This is not a root-cause fix. The proposal, test, and spec edit are archived in rolling-wheel/rejected-source-lift-order-1; none remains in production.

The latter experiment also disproves an expression-size explanation: the two orientation residual programs grew from 5,002 to 6,516 operations apiece. The changed trajectory’s success cannot establish a general numerical repair. revolute-point-replay-1.log records both representations and both saved directions through the public interpreter runtime. OMC’s saved _04set.c uses two dynamic state sets, each choosing one of three candidates, besides the joint’s two states. This explains a substantial representation difference; it does not by itself prove the needed Rumoca transformation.

The standalone source fixture crates/rumoca/tests/fixtures/index_reduction/RateCancellation.mo removes MSL and retains the essential computation: three Euler angles, an implicit angular rate map, two rotation constraints, and two unknown angular accelerations. At q={0,-acos(epsilon),0}, w={0,-2.4010301622267085,0}, the directional runtime probe in the w[1] direction produces:

Differentiating the two rotation constraints twice at q[1]=q[3]=0 gives der(w[1])=-w[2]*w[3] and der(w[3])=w[2]*w[1] for the constructed RHS. Consequently its w[1] partial is w[2], independent of the nonzero cosine. The zero result is an inaccurate accepted numerical derivative, not merely an overly strict residual check. This probe concerns the constructed RHS Jacobian, including directions normal to the retained manifold; it is not a new claim of a completed wrong MSL trajectory. A simpler linear two-coordinate rate map remained accurate and is preserved as a rejected reduced reproduction.

Evidence lives in rolling-wheel/revolute-jvp-registers-1.json, revolute-source-lift-structure-1.json, revolute-point-replay-1.log, and implicit-rate-cancellation-probe-{1,2,3}.{rs,log}. The last probe removes the tiny off-axis angle offsets and retains the failure. Temporary test code has been removed. The next repair must preserve the coupled implicit rate equations or otherwise establish stable derivative evaluation; lifting the observation alone is insufficient. This remains a hypothesis to prove with the reduced fixture and saved real-model point before another full-cohort run. No baseline, exclusion, tolerance, canary list, or cohort parity claim changed.

Repair the invariant-value preflight/reconstruction disagreement

The full sweep’s ConstantActuator panic reproduced in a model with offset=p+p; x*x=offset; der(x)=velocity. The equality closure can prove that offset is invariant without assigning a single expression as its payload. Value preflight then used the exact causal definition, but reconstruction selected the payload-less equality and panicked. Both now select their exact expression and sign through one shared helper. This preserves the distinction between a known derivative and a known value required by STRUCT-T03.

The reduced test failed with the same panic before this repair. Positive and negative invariant definitions now simulate correctly with both integrators, checking every output channel. ConstantActuator’s ordinary worker now returns EL005 (77/82 matched), with no panic or simulation credit. Its structural stage took about 68 ms in the artifact-producing diagnostic attempt; the initially observed wall delay was outside that stage. This is not a new cohort claim.

All 1,071 focused tests pass (99 library, 625 core, 141 Solve, 206 structural), along with phase/core Clippy and workspace formatting. Tier 1 multibody-invariant-value-{frontier,origin,canary} has no phase or band changes against multibody-single-anchor-*: 5/6/9 compared high, 1/2/0 reviewed exclusions, zero missing/nonidentifiable traces. Evidence is in rolling-wheel/invariant-value-*.

RevoluteConstraint’s EX002 sensitivity refusal independently reproduces with the same row and residual. Repairing that lost high band remains the next priority; the earlier PrismaticConstraint investigation is paused.

Full sweep finds UniversalConstraint progress and a Revolute regression

target/msl/multibody-single-anchor-full, at 2da150a3f870d756ecc5f1242fbc36f09adc17b7, completes the 566-model gate. It has 167 strict-high models, 167 compared, 20 reviewed exclusions, and zero missing or nonidentifiable traces. MultiBody remains 25/42 high, with 25 compared and the existing BevelGear1D exclusion. These unchanged totals hide an important exchange: UniversalConstraint becomes high, while previously high RevoluteConstraint now refuses during runtime sensitivity validation. The repository baseline gate passes, but this branch-to-branch regression must be repaired before further breadth work or a PR.

UniversalConstraint has 1,079 high and one minor trajectory channel, with no deviating or severe channels. RevoluteConstraint initializes, then reports EX002: algebraic projection sensitivity residual 1037, targeting bodyOfConstraint.body.z_a[3], has magnitude 1.164153e-10 against its unchanged 1e-10 threshold. This is a runtime refusal, not a completed wrong trace; its former high band is nevertheless lost. All 24,935 compared initialization channels are high, and no completed model has a deviation channel.

The sweep also exposes a constructor/preflight panic in ConstantActuator (holonomic value preflight proves a materializable anchor) where there was previously an explicit structural refusal. Engine1b now initializes but still times out. Inverse_sh_TX reaches DAE construction instead of failing Flatten. These changes and all artifacts are recorded in rolling-wheel/single-anchor-full-delta-1.json. The tracked tree was clean; the run’s dirty marker includes the preserved comm_fastdyn.md. Baseline promotion, combined quick/full release verification, push, and PR remain pending.

Independent constraints need no minimum state-declaration count

The exact PrismaticConstraint orientation residual passed first-derivative and retained-value preflight. A temporary probe ruled out missing function argument values: both matrices and angular velocities were reconstructible. It instead found one state anchor and an algebraic read, which an old guard rejected for having fewer than two distinct state declarations. One rotation matrix is a tensor declaration; its declaration count says nothing about whether the orientation constraint is independent.

STRUCT-T03 admission now relies on the existing source-value identity and self-definition exclusions, exact derivative/value preflight, and nonempty state dependencies. The declaration-count heuristic and its bookkeeping are removed. This does not loosen matching or initialization checks. A reduced structural test for observed=x+x; observed*observed=1 failed before the change and now admits only the independent constraint, for both scalar and vector states. The companion test continues to reject the observation’s defining equation. An end-to-end angular-observation test already passed through another reduction route before this repair; it remains a regression, not the red proof. Both integrators reproduce all seven analytical channels, and the ordinary worker agrees highly with OMC on all seven shared trajectory and initial channels. A diagnostic receipt script stopped on a Python name typo after comparison; its continuation validated the saved traces without rerunning them.

All 1,070 focused tests pass (99 library, 624 core, 141 Solve, 206 structural), as do phase/core Clippy and workspace formatting. Tier 1 multibody-single-anchor-{frontier,origin,canary} has no phase or band changes against multibody-atan2-constraints-*: 5/6/9 compared high, 1/2/0 reviewed exclusions, and zero missing or nonidentifiable traces. No temporary production probe remains. Evidence is in rolling-wheel/single-anchor-* and rolling-wheel/prismatic-orientation-preflight-*.

PrismaticConstraint’s orientation owner 450 now appears in the ordinary reduction candidates and is selected. The real model still refuses preparation; the last accepted intermediate remains 2,314/2,317 matched, with the same three unmatched matrix-equality rows. The quaternion-rate reconstruction and the order of the retained constraints are the next investigation, not an established root cause. No additional MultiBody parity credit is claimed.

Second derivatives of angular constraints

OMC differentiates PrismaticConstraint’s Orientation.equalityConstraint twice: its generated equations 1328–1330 retain the source Prismatic joint’s line 55 and second derivatives of the rotation matrices. Rumoca’s structural builtin profile admitted atan2 only at first order. Under STRUCT-T03 and the MLS function-differentiation contract, it now constructs the second derivative with D=x*x+y*y, N=x*y'-y*x', and theta''=(x*y''-y*x''-2*theta'*(x*x'+y*y'))/D. The original constraint and denominator remain; no domain cancellation or scalar tensor expansion occurs.

The reduced Atan2Constraint fixture was structurally singular (4/5 matched) before this change. Both integrators now reproduce all five analytical channels for constant and changing first arguments. A zero-radius initial orientation remains rejected. The normal worker and OMC agree highly on all five trajectory and initialization channels over the full 0–0.1 fixture interval. The initial OMC probe with nonzero radial velocity instead violated its requested initial velocity; that oracle defect is preserved separately in prismatic-atan2-omc-initial-gap-1.json. The nonzero-velocity case remains an analytical regression, and no comparator exclusion was added.

All 1,068 focused tests pass (99 library, 623 core, 141 Solve, 205 structural), along with phase and core Clippy and workspace formatting. Tier 1 multibody-atan2-constraints-{frontier,origin,canary} has no phase or band changes against multibody-initial-obligations-*: 5/6/9 compared high, 1/2/0 reviewed exclusions, and zero missing or nonidentifiable traces. The canary’s eleven existing failures remain visible. Evidence is in rolling-wheel/prismatic-atan2-* and rolling-wheel/atan2-constraints-*.

This fixes a demonstrated prerequisite, but the real PrismaticConstraint still refuses structural preparation. Its inspected intermediate remains 2,314/2,317 matched, and the orientation residual (owner 450) never becomes a candidate. The next investigation checks the distinct derivative and retained-value prerequisites for that exact source equation. No new MultiBody coverage or release-readiness claim follows from the reduced-model result.

Full cohort preserves all high bands after initialization repair

target/msl/multibody-initial-obligations-full passes at f4ec3b36241cafdca93c2d18af935c05df580400: 167/566 strict-high (29.51%), 167 compared, 20 reviewed exclusions, and zero missing or nonidentifiable traces. MultiBody remains 25/42 high, with 25 compared and one reviewed BevelGear1D exclusion. There are no changed model bands and no actionable counterexamples. All 24,773 initialization channels are high; trajectory channels total 24,727 high, 46 minor, and zero deviating or severe. MovingCoilActuator.ArmatureStroke has one minor channel instead of zero, while retaining its high model band. This is recorded rather than hidden by the unchanged model count.

The sole phase/status delta is DC_CompareCharacteristics: its Solve timeout becomes an explicit structural refusal (EL005). It still does not simulate and gains no parity credit. The compiler tree is committed; the dirty marker includes the preserved untracked comm_fastdyn.md. The complete receipt is rolling-wheel/initial-obligations-full-delta-1.json. Combined quick/full release verification, baseline promotion, push, and PR remain pending.

Fixed initialization equations survive algebraic demotion

PrismaticConstraint exposed an obsolete structural preservation check. Its FreeMotionScalarInit component fixes two position and two velocity outputs. The position outputs alias the scalar derivative-block inputs. Directly reconstructing those inputs from the body’s position tensor was refused because the equality class would no longer contain a retained state. OMC’s generated initialization code preserves those starts explicitly. Rumoca’s Solve phase also already emits every fixed continuous Real equation, independently of its state or algebraic role, through transferred rows and its remaining-coordinate inventory. The structural check was still using the older state-seeding model.

Preservation now inventories the source-fixed declarations themselves. The existing checked reconstruction carries their exact fixed attributes and mapped start expressions; Solve retains every scalar equation, including default starts and coordinates not covered by a transfer. Equality-class transfer remains an equivalent representation, not permission to omit other initial equations. This follows MLS section 8.6 and the initial-equation construction contract in SPEC_0043. No solver tolerances, source equations, state preferences, or model-name conditions changed.

The reduced ProjectedInitialCoordinate model has a two-component state x, observation u=x[1]+2*x[2], der(u)=rate, and fixed scalar observations pin=u=1 and second=x[2]=2 at initialization. Before the repair its trajectory was correct but it retained three scalar states; the state-count regression failed. It now retains only x and both solvers reproduce the exact exponential solution. The ordinary worker and OMC agree highly on all six shared channels over the complete 0–1 interval, with analytical errors below 1e-6. The first diagnostic driver mistakenly requested OMC only through 0.1 and stopped before comparison; the corrected OMC interval reuses the successful full Rumoca trace. Conflicting fixed states, time constraints, and constant constraints still fail initialization. Structural tests check the retained original continuous equation and fixed initial equation after direct demotion.

All 1,065 focused tests pass: 99 library, 620 core, 141 Solve, and 205 structural. Phase all-target/all-feature Clippy, core-suite Clippy, and workspace formatting pass. multibody-initial-obligations-{frontier,origin,canary} has empty phase and band deltas against multibody-early-tensor-rate-*: 5/6/9 compared high, 1/2/0 reviewed exclusions, and no missing or nonidentifiable traces. The fixed canary’s eleven existing failures remain visible.

PrismaticConstraint is still a structural refusal and gains no parity credit. Its source DAE is identical before and after the change (558eab4e8f4d9219263cd690753b909507b2956e0a5cf620ff0086e75c7db8e9). The current ordinary worker reports 2,287/2,313 matched at the final refusal; the public inspector’s last accepted intermediate instead has 2,314/2,317 matched, the same three unmatched rows as before. That distinction prevents the final error from being mistaken for the intermediate being investigated. Rows 2,289–2,291 are the last three components of the exact connection equality world.frame_b.R.T = fixedRotation.frame_a.R.T. The intermediate retains the constrained body’s quaternion and several orientation matrices as states. This localizes the next comparison; it does not yet prove why those rows cannot be matched or establish that their source equation is wrong.

Evidence is in rolling-wheel/prismatic-initial-*, rolling-wheel/prismatic-preserved-initial-*, rolling-wheel/prismatic-stalled-rows-2.log, and rolling-wheel/initial-obligations-*. The complete cohort sweep follows this milestone; no new cohort total or release readiness is claimed here.

Full sweep confirms 25 high-parity MultiBody examples

target/msl/multibody-early-tensor-rate-full passes the complete 566-model gate at e1aacdd07518f988be542e0b29480992206113dc. There are 167 strict-high models (29.51%), with 167 compared, 20 reviewed exclusions, zero missing traces, and zero nonidentifiable traces. The tracked compiler tree is clean; the run’s dirty marker includes the preserved untracked comm_fastdyn.md. Of 24,773 trajectory channels, 24,728 are high and 45 have minor differences; none are deviating or severe. All 24,773 initialization channels are high. Raw simulation completion is 187 and receives no additional parity credit.

MultiBody has 25 of 42 strict-high examples, with 25 compared, one reviewed BevelGear1D exclusion, and no missing or nonidentifiable traces. Compared with multibody-projected-initial-full at c74dd524, RevoluteConstraint is the sole band change: absent to high. It completes its full experiment with all 918 shared trajectory and initialization channels high. Solve preparation takes 7.173 seconds, including 0.162 seconds in structural reduction; simulation takes 2.076 seconds. No previously high model regresses and no new actionable trace counterexample appears.

The remaining sixteen MultiBody failures are seven Solve timeouts, two structural refusals, two Flatten timeouts, three record-input derivative specialization refusals, one function shape-proof refusal, and one unsupported array-valued fixed attribute. The seventeenth non-high model is the reviewed BevelGear1D exclusion. These outcomes receive no high-parity credit.

One other phase outcome changes: Inverse_sh_TX again reports the missing MixtureGasNasa.ThermodynamicState constructor layout in Flatten (EF015) instead of the prior ToDae Array/Enumeration mismatch (ED019). This previously observed diagnostic variation remains unexplained; neither result simulates, and the intervening fixes do not modify Flatten or earlier phases.

The full log and complete phase/band delta are rolling-wheel/early-tensor-rate-full-{1.log,delta-1.json}. The evidence receipt binds the implementation, normal workers, focused checks, and cohort artifacts. Combined quick/full release verification, baseline promotion, push, and PR remain pending. The 100% MultiBody goal is not complete.

RevoluteConstraint completes with high OMC trace agreement

The ordinary worker now compiles, initializes, and simulates Modelica.Mechanics.MultiBody.Examples.Constraints.RevoluteConstraint over its complete 0–10 second interval. All 918 shared OMC trajectory channels and all 918 initialization channels are high agreement, with zero minor, deviating, or severe channels. This is a focused model result; the next full-cohort run must establish the new MultiBody count and detect any other changes.

The original DAE is byte-identical before and after the repair (16554864b9cc638195b8f7f9c0ff66a1199ff0c220c694bcfd4f75c3f057e45d). The first divergence was structural preparation: initAngle.angle has an implicit coupled angular-rate equation, but its derivative alias was introduced only after complete structural matching. Index reduction needed that rate representation to differentiate the rotational constraints a second time. The source-owned angular-rate relation is the same one used by OMC’s generated equations; no model-specific equations or state choices were introduced.

Structural preparation now exposes implicit tensor rates before holonomic reduction through the existing checked whole-tensor alias reconstruction. Actual partial matching proves complete coordinate coverage of explicit derivative owners. It cannot substitute for complete structural analysis or certify a coupled native derivative block for this earlier phase. Complete prepared systems retain their native derivative blocks. The real-model probe showed three native-block matches but no explicit derivative definition for the angle tensor; after reduction even those native-block matches disappeared. That distinction explains why moving the previous selection unchanged was insufficient. STRUCT-T09 records the sequencing and owner-coverage contract; MLS Appendix B and the exact equality der(state)=alias establish equivalence.

The new representation can expose direct state reductions, so those are exhausted before differentiating a constraint. The reduced implicit-rate example consequently retains the same four scalar states as its explicit-rate form instead of integrating a redundant velocity state. Scalar, tensor, and domain derivative owners keep their existing checks; aliases retain complete shapes and introduce no independent initial conditions. No incomplete matching escapes as a prepared proof, no tensor basis is enumerated, and no tolerances, attempt limits, or model-name branches changed.

The regressions cover diagonal and coupled tensor rates, an independent constraint requiring a second derivative, explicit domain owners, and an already complete coupled ODE that must retain its native derivative block. Both integrators match analytical trajectories. A split scalar-projection variant still has a separate reconstruction limitation: its test checks that partial explicit coverage receives the required whole-tensor alias, not that the model simulates. It remains a visible structural refusal.

All 1,062 tests pass: 99 library, 617 core, 141 Solve, and 205 structural. Phase all-target/all-feature Clippy, core-suite Clippy, and workspace formatting pass. The two ordinary-worker reduced models each compare 7/7 shared OMC channels high and satisfy analytical checks. The fixed Tier 1 targets multibody-early-tensor-rate-{frontier,origin,canary} have empty phase and band deltas against multibody-affine-map-complete-*: 5/6/9 compared high, 1/2/0 reviewed exclusions, and zero missing or nonidentifiable traces. The canary’s eleven failing models remain visible.

Evidence is in rolling-wheel/early-tensor-rate-*, rolling-wheel/revolute-early-tensor-rate-worker-2, and rolling-wheel/revolute-early-tensor-rate-comparison-2.json. The source-selection probe was removed. The full-cohort sweep is next; combined quick/full release verification, baseline promotion, push, and PR remain pending.

Affine tensor reconstruction retains component proofs

The saved RevoluteConstraint angular-rate equations expose two missing proof steps. Frames.absoluteRotation combines a transformed angular velocity with an additive relative velocity, so its source map is A*q+c=b. The existing tensor reconstruction rejected any nonzero c. The reconstruction now retains b-c as its right-hand side and includes the offset’s state dependencies. The coefficient and offset must still be independent of the reconstructed unknown; nonlinear coefficients remain refused, and numerical singularity remains the checked aggregate solve’s responsibility. Governing contracts are MLS sections 8.3.1 and 10.6, Appendix B, SPEC_0007 STRUCT-T03’s square linear reconstruction and source-projection requirements, and SPEC_0032’s compact tensor ownership.

The next source map depends on a vector proved by an earlier map. Reconstruction facts now extend to a fixed point, admitting each aggregate and component proof once. Both kinds must advance together. An aggregate-only iteration initially regressed six existing contact tests: after delta was independently proved, road=delta+r gained an auxiliary owner that hid the literal-array projection needed to prove z. The diagnostic observed that component proof after the first layer and its absence after the second. Retaining the component proof before extending the next aggregate layer restores all six tests. Existing proofs are never replaced; termination follows from the finite variable catalog. Source equations, state roles, assertions, and dependency checks are preserved, without a depth cutoff or new scalar tensor owner.

The focused regressions check offset dependencies, nonlinear refusal, equal proof size at vector extents 3 and 4096, and two- and four-stage reconstruction chains, including reverse source order. Both BDF and RkLike preserve the analytical position, velocity, acceleration, and auxiliary values. The two ordinary-worker fixtures compare all 6 and 8 shared OMC channels high, with zero minor, deviating, or severe channels. Both compilers also satisfy the analytical checks. All 957 tests pass: 205 structural, 141 Solve, and 611 core. Both phase crates’ all-target/all-feature Clippy, core-suite Clippy, and workspace formatting pass.

multibody-affine-map-complete-{frontier,origin,canary} records empty phase and agreement-band deltas against the corresponding multibody-indexed-parameter-guard-* runs: 5/6/9 compared high, 1/2/0 reviewed exclusions, and zero missing or nonidentifiable traces. The fixed canary retains its eleven failing models. These are Tier 1 regression checks, not a new full-cohort measurement or MultiBody coverage gain.

The diagnostic RevoluteConstraint DAE with the previously selected derivative alias now reaches complete structural matching: its two rotational constraints reduce the deficit from four to two to zero. This is not an ordinary simulation pass. The normal worker still refuses Solve with EL005, reporting 2,181/2,208 matching and no trace. The next task is to make the needed implicit derivative representation available to index reduction through the normal compiler path. A reduced model with 2*der(q)+q=omega, der(omega)={0,0}, z=q[1], der(z)=v, and der(v)=force reproduces a 6/7 structural refusal; OMC simulates it and matches all seven analytical observables within 1e-8.

Making the equivalent derivative alias explicit in that Modelica source (Real rate[2]; der(q)=rate; 2*rate+q=omega) lets the ordinary Rumoca worker simulate the same constrained motion. All nine shared OMC channels are high, and both compilers match all nine analytical observables within 1e-6. OMC’s generated C supplies der(q)=0.5*(omega-q), v=der(q[1]), and force=0.5*(der(omega[1])-der(q[1])). This isolates the missing implicit-rate representation needed by the second derivative; it is not evidence that ordinary implicit-rate compilation or RevoluteConstraint is repaired.

Evidence is under rolling-wheel/affine-map-*, rolling-wheel/revolute-affine-map-complete-*, and rolling-wheel/implicit-rate-frontier-1. Temporary probes were removed. No full-cohort sweep, combined quick/full release gate, baseline promotion, push, or PR was performed for this change; 100% MultiBody remains unproved.

Parameter-array guards retain their function call context

RevoluteConstraint’s stalled structural DAE discovers no further candidates. Tracing source owners 332 and 333 (0=ex_a*R_rel.T*e and 0=ey_a*R_rel.T*e) proves that their first differentiability check fails before the state-anchor count check. The failing expression is axisRotation’s conditional guard: its axis argument selects an element of initAngle.sequence_start through axesRotations. The array is a parameter, but the invariant-guard proof did not handle indexing. OMC’s retained generated _03lsy.c differentiates the same rotational constraints twice.

The structural proof now follows both the array and its subscripts through the exact function call context. Both must be invariant. A reduced nested-function regression first reproduced a 4/5 singular system, then exposed an additional reconstruction error: the array argument was substituted while the subscript still referenced the callee’s parameter. Instantiated value and derivative reconstruction now instantiate subscripts through the same context, sharing the existing subscript reconstruction implementation. Ordinary function-body replay retains its original owner. The governing contracts are MLS sections 3.6.5, 3.8.3, 10.5, and 12.4, SPEC_0007 STRUCT-T03’s exact function substitution, and SPEC_0036 / SPEC_0043’s expression ownership requirements.

Both branch choices pass analytical position, velocity, and acceleration checks on BDF and RkLike. Varying array contents and varying indices remain refused. The ordinary worker’s two reduced traces each have all five shared OMC channels high, with no minor, deviating, or severe channels. The 954 tests pass (204 structural, 141 Solve, 609 core), as do both phase crates’ all-target, all-feature Clippy, core-suite Clippy, and workspace format checks. multibody-indexed-parameter-guard-{frontier,origin,canary} records empty phase and agreement-band deltas against the corresponding multibody-projected-initial-complete-* targets: 5/6/9 compared high, 1/2/0 reviewed exclusions, and zero missing or nonidentifiable traces. The fixed 20-model canary still contains its eleven previously failing models. These are Tier 1 checks, not a new full-cohort measurement.

The unchanged RevoluteConstraint source DAE is byte-identical before and after the fix. Replaying its closest stalled snapshot now discovers three initial lift candidates and reaches the two rotational constraints, retaining four manifold roots. It still matches only 2,206 of 2,208 equations. The normal model attempt also still refuses Solve with EL005 (2,181/2,208 in its final original- system report), so it receives no simulation or coverage credit. The retained rotational rows contain first derivatives; identifying the exact proof that prevents the next reduction, against OMC’s second derivatives, remains open.

Evidence is under rolling-wheel/indexed-parameter-guard-*, rolling-wheel/revolute-preflight-probe-*, and rolling-wheel/revolute-indexed-guard-*. Temporary probes were removed. No combined quick/full release gate, full-cohort rerun, baseline promotion, push, or PR was performed for this change.

Remaining angular-rate definition

At implementation 3355ea28, a second diagnostic proves both rotational constraints now pass first differentiation and value materialization, but fail second differentiation and retained-tangent materialization. The first rejected second-order leaf is state 996, freeMotionScalarInit.initAngle.angle; its explicit derivative-definition entry is absent. The other retained states have explicit derivative owners. Thus the parameter guard is repaired, while the angle-rate relation needed for another derivative remains unavailable.

A diagnostic application of the existing whole-tensor derivative alias adds der(angle)=alias and preserves all source equations. This alone still fails: 2,209 of 2,211 equations match, retaining the same two-equation deficit. It is not a production change or a coverage gain. The next investigation must trace the exact axesRotations angular-velocity equation and its independent value anchors, rather than assume that moving normalization earlier suffices. revolute-derivative-order-probe-* and revolute-derivative-alias-inspection-1 retain this evidence. Both temporary diagnostic tests completed and were removed; the 954-test verification belongs to the committed production repair, not those probes.

Full cohort after projected initial-value transfer

target/msl/multibody-projected-initial-full passes the complete 566-model gate at c74dd524fb89a811ef6a8813fbb21f4759ef1f4f. There are 166 strict-high models (29.33%), with 166 compared, 20 reviewed exclusions, zero missing traces, and zero nonidentifiable traces. The 23,855 trajectory channels comprise 23,810 high and 45 minor channels, with zero deviating or severe channels; all initialization channels are high. MultiBody remains 24/42 high, with its reviewed BevelGear1D exclusion receiving no high credit. No model changes its agreement band from multibody-scalar-state-offset-full at 2f3e8ee2.

The only phase change is Inverse_sh_TX returning from Flatten EF015 to its earlier ToDae ED019 Array/Enumeration rejection. It still does not simulate; the cause of this diagnostic variation remains unproven. RevoluteConstraint still fails structural analysis under the normal budget, after accepting the previously blocked projected-state demotion. Neither case is a coverage gain. The full gate and complete delta are rolling-wheel/projected-initial-full-*. No baseline promotion, combined quick/full release gate, or PR has been performed.

Preserve fixed starts through tensor projections

RevoluteConstraint reaches ToDae, then structural reduction refuses to demote freeMotionScalarInit.derd[2].u because it would discard the fixed start on freeMotionScalarInit.angle_2. The output aliases that scalar state, which aliases element two of the retained initAngle.angle vector. The first executed ProjectedPin regression reproduces WouldDiscardInitial for output q and retains both states. The equality closure previously recognized whole variables and singleton projections, but not components of larger arrays.

Structural initial-value transfer now gives authored literal projections sparse component identities in its existing signed equality closure. The same facts supply demotion preservation and transferred initialization rows. Dynamic indices and nonlinear dependencies cannot establish a value alias. Array declarations, source equations, and provenance remain intact; no tensor basis is enumerated. The governing contracts are MLS sections 8.6 and 10, SPEC_0007 structural lowering / STRUCT-T03, SPEC_0036 Expressions and Equations, and SPEC_0043 section 4.

The existing OMC reference tool completes the unchanged RevoluteConstraint to ten seconds. Generated initialization equation 970 preserves the second angle’s fixed start while the adjacent nonlinear initial system solves the other angles. The projection candidate accepts state 999’s formerly blocked demotion and emits no WouldDiscardInitial records. The model still cannot simulate: the closest retained snapshot matches 2,206 of 2,208 equations, leaving f_x[2200], f_x[2201], constraint.P, and fixedRotation.frame_b.t[1] unmatched. Its final failure reports the original singular system. This is no coverage gain and RevoluteConstraint parity remains unmeasured. revolute-projected-initial-* retains the diagnostic and public structural inspection. Diagnostic worker 1 was stale and is marked invalid; worker 2 contains the projection change.

Review also exposed and closed a counterexample in the uncommitted candidate. PartialPin declares a[3](start={1,2,3},each fixed=true), defines a={x,0,0}, and separately references only a[1]. The candidate incorrectly completed with initial a={1,0,0}: Solve treated one transferred component as covering the entire source declaration. Each transferred pin now carries its exact source scalar, and Solve retains every uncovered component equation. The saved Solve artifacts show one initial residual becoming all three required residuals. The normal worker now refuses initialization with residual -3; OMC independently rejects the same source for $START.a[3]-a[3] = 3. This closes the counterexample by rejecting an inconsistent model. Regression variants select each of the three components and test both inconsistent and consistent systems on BDF and RkLike. Evidence is in projected-initial-partial-source-1, including the original wrong trace, source, binary hashes, initial-row delta, and closure receipt.

Both valid projection fixtures retain one three-component state declaration and simulate the full one-second interval. Each compares all six shared OMC channels high, with no minor, deviating, or severe channels. Both implementations agree with the analytical exponential solution within 1e-5. Structural tests cover signed aliases, offsets, vectors through 4,096 elements, matrices, distinct indices, fixed targets, and refusal to infer identity from dynamic/nonlinear reads. All 204 structural tests, 141 Solve tests, and 607 core integration tests pass; both phase crates pass all-target, all-feature Clippy.

The final Tier 1 runs are multibody-projected-initial-complete-{frontier,origin,canary}. They retain five, six, and nine high comparisons respectively, with one, two, and zero reviewed exclusions and no missing traces. Every phase and band delta against the prior fixed target sets is empty. Receipts are projected-initial-complete-*-delta-1.json; the complete worker build manifest binds both the normal executables and MSL test binary. These focused results do not establish a new cohort number. No baseline promotion or PR is claimed.

A separate Flat inspection establishes that RollingWheelSetPulling’s reported “record input” derivative failure actually names the external-object table handle passed to getTimeTableValueNoDer2. Record decomposition already ran; that path needs a checked external-object lifecycle. wheelset-derivative-inputs-* retains the evidence. No change or coverage gain is claimed for that finding.

Full cohort after constant-offset state reduction

target/msl/multibody-scalar-state-offset-full passes the complete 566-model gate at 2f3e8ee21157be462ed67e11dfa6cbd5ef44e492. The compiler tree is committed; the run’s dirty marker includes the preserved untracked comm_fastdyn.md. There are 166 strict-high models (29.33%), with 166 compared, 20 reviewed exclusions, zero missing traces, and zero nonidentifiable traces. Of 23,855 trajectory channels, 23,810 are high and 45 have minor differences; none are deviating or severe. All 23,855 initialization channels are high. Raw simulation completion is 186 and receives no additional parity credit.

MultiBody now has 24 of 42 strict-high examples. Its reviewed BevelGear1D exclusion remains outside that count. LineForceWithTwoMasses is the sole band change from multibody-lift-owners-full-11 at 6da1462a: absent to high under the normal budgets. Its previously recorded force counterexample is closed by both the diagnostic and ordinary full-cohort comparisons. No previously high model loses its band, and no new actionable trace counterexample appears.

One other phase outcome changes and remains unresolved: Modelica.Media.Examples.SolveOneNonlinearEquation.Inverse_sh_TX previously reached ToDae and rejected parameter s_min with an Array/Enumeration type mismatch (ED019). It now stops in Flatten because the exact constructor record layout for MixtureGasNasa.ThermodynamicState is missing (EF015). It produces no simulation in either run. The commits between these complete sweeps change DAE and structural code, but no Flatten or earlier-phase source; this observation does not establish the cause of the changed diagnostic. It remains an explicit triage item, not a claimed fix or a trace counterexample.

rolling-wheel/scalar-state-offset-full-{1.log,delta-1.json} retains the full result and the complete phase/band delta. No baseline promotion, combined quick/full verification, or PR has been performed. The 100% MultiBody goal remains open.

Constant offsets admit derivative-only state definitions

The remaining damper1.s_rel in LineForceWithTwoMasses reduces to the source relations length=q+offset, length=sqrt(x*x+1), der(x)=-x, and der(q)=rate. With a constant parameter offset, Rumoca initially retains both x and q. Writing the equivalent equation q=sqrt(x*x+1)-offset already leaves only x. This rejects nonlinear differentiation as the explanation: the first divergence is structural discovery ignoring its existing proof that der(length)=der(q). OMC backend output for both exact scalar sources retains only x and computes q=length-offset and rate=x*der(x)/length. Both OMC simulations complete. The contract is MLS section 8.3.1 and Appendix B, within SPEC_0007 structural lowering and SPEC_0040 STRUCT-T03.

Discovery now tries exact value anchors before signed constant-displacement anchors. A distinct derivative-only definition permits replacing der(q) while preserving every original value equation and initialization obligation. It cannot substitute q into a retained manifold: equal derivatives do not prove equal values. The regression covers both offset signs, scalar, singleton, three-element and 4,096-element vectors, and a matrix. Dependent declarations are eliminated without scalarizing them; vector expression count stays constant. A time-varying offset supplies no derivative equality, and a manifold requiring the displaced state’s value rejects the derivative-only definition.

The singleton case also exposed a stale anchor index that kept only scalar projections of singleton state arrays. Whole coordinates and existing projected views are now indexed separately. Direct substitution requires the matching shape; algebraic payload substitution retains its checked singleton reshape. An existing scalar-to-array projection regression caught an overly restrictive first candidate, and passes after that distinction. All 200 structural tests and all-target, all-feature structural Clippy pass. Evidence is under rolling-wheel/scalar-state-offset-*.

The artifact-enabled LineForceWithTwoMasses diagnostic now retains exactly revolute1.phi, revolute1.w, revolute2.phi, and revolute2.w: four scalar states, matching OMC, instead of the previous 62-state fallback. Initialization passes and the full three-second trajectory completes in 501 samples. The canonical comparator, linked to the current worker’s libraries, reports all 879 trajectory channels and all 879 initialization channels high, with zero minor, deviating, or severe channels. The maximum force-difference magnitude falls from approximately 1.015e-4 to 1.42e-13. This closes the four-channel refinement counterexample with a general structural repair and focused regressions, without changing solver settings or comparator tolerances.

The diagnostic records 0.189 seconds for integration, but structural preparation and Solve lowering together take 10.434 seconds. This artifact-enabled run earns no ordinary-budget coverage or performance credit. Its optional Modelica debug render also reports unsupported linear_solve; the DAE and Solve JSON artifacts and complete simulation trace are retained. The unchanged originating model remains in the ordinary focused gate and the next full cohort comparison.

Both exact scalar fixture sources also run through Rumoca’s complete compiler and simulator with one state. Their x, length, q, and rate traces agree with the analytical solution, and the canonical comparator compares every available shared OMC channel high over their common quarter-second interval. The initial fixture request compiled only because it omitted the explicit non-MSL simulation-target flag; the second request supplies that flag and is the first simulation attempt. The first diagnostic driver similarly failed before launching Python, and the absolute interpreter launched the sole model attempt. Receipts include scalar-state-offset-{numeric,comparator,diagnostic}-*.

Tier 1 is complete. The unchanged six-model frontier list now admits LineForceWithTwoMasses under the ordinary budgets and compares its complete trace high. The five compared models are high, with one reviewed exclusion and no missing or nonidentifiable traces. The ordinary LineForceWithTwoMasses row records 6.735 seconds for structural preparation plus Solve lowering and 0.0985 seconds for integration. The unchanged eight-model origin list retains all six high comparisons and its two reviewed exclusions. The fixed twenty-model canary retains its nine high comparisons and all prior phase outcomes, with no exclusions. Every band or execution change is the originating model’s improvement; all other deltas are empty. Receipts are scalar-state-offset-*-delta-1.json. Workspace formatting and whitespace checks pass. A new full cohort comparison at the committed checkpoint is next; no cohort gain, baseline promotion, complete quick/full gate, or PR is claimed from these Tier 1 results.

Supplied derivative chains survive successive state demotions

The reduced LineForceWithTwoMasses path exposed a lost MLS section 12.7.1 predecessor. Structural differentiation emitted ordinary calls to maxWithoutEvent_d, discarding the fact that they came from differentiating maxWithoutEvent. A later reduction round consequently could not select its derivative(order=2) annotation. The functions and annotations were present; the call’s differentiation history was missing.

DAE calls now retain their source call and selected derivative link. Checked construction and current-wire replay prove the predecessor, unchanged argument prefix, and tangent signature. Schema 38 replaces schema 37 without a compatibility reader. Reconstruction maps derivative identities explicitly and replays links as their function components become available, including calls inside function bodies. Value reconstruction preserves the original call and shared result projections. Ordinary calls to a derivative function acquire no history. The owning contract is SPEC_0036’s function-derivative proposal and SPEC_0043 section 10, within SPEC_0007’s structural reconstruction contract.

A compact regression defines q=f(s)*ones(n), der(q)=w, and der(w)=a, with independent states s,v and a conditional function supplied with first and second derivatives. Disabling only predecessor-aware selection retains both dependent state arrays; restoring it leaves s,v, with no manifold and constant IR size at array lengths 1, 3, and 4,096. A function-body wrapper preserves both links across reconstruction. DAE tests reject forged links, prefixes, and origins, and round-trip JSON/bincode with compact tensors and multiple result projections. All 190 DAE tests, 196 structural tests, and both crates’ all-target, all-feature Clippy pass. OMC independently emits q=f(s), w=fd(s,v), and a=fdd(s,v,der(v)) for the three-element fixture; its quarter-second simulation completes and agrees with the analytical oscillator solution within 2.40e-6.

The actual intermediate LineForceWithTwoMasses DAE improves from eleven to five scalar states: both three-element v_CM state arrays are eliminated. Its unmatched equations and unknowns each decrease from seven to one. The remaining state declarations are the two revolute angles and velocities plus damper1.s_rel. The last unmatched position row is f_x[2161], within owner 686’s rod1.frame_b.r_0-jointUPS.frame_a.r_0; the unmatched unknown is jointUPS.f_bd_a[3]. OMC backend equation 123 defines damper1.s_rel=jointUPS.axisLength-jointUPS.s_offset, and equation 84 computes axis length from the position norm. Rumoca retains both source relations, but the direct-state detector finds only aliases anchored on the same length state. Independent affine-offset reconstruction is the next hypothesis; it has not yet been isolated in a minimal regression or repaired.

The reduced system still fails complete matching, so the final prepared model falls back to 62 scalar states. The artifact-enabled diagnostic initializes but hits the unchanged 12-second simulation limit. It earns no coverage or timing credit and does not close the earlier four-force-channel counterexample. All temporary probes are removed. Evidence is under rolling-wheel/derivative-chain-*, line-force-derivative-chain-*, and line-force-last-*.

Tier 1 multibody-derivative-chain-{frontier,origin} and multibody-derivative-chain-canary-2 retain all prior bands and phase/execution statuses on the unchanged six-, eight-, and twenty-model lists. Their four, six, and nine comparisons are all high, with one, two, and zero reviewed exclusions, and no missing or nonidentifiable traces. The original canary output name already existed, so the driver stopped before launching that gate; the fresh canary-2 directory contains its sole new attempt. The ordinary LineForceWithTwoMasses attempt still exceeds the Solve budget. No new cohort claim, baseline promotion, complete quick/full verification, or PR is claimed.

Accepted-step projection amplifies the LineForceWithTwoMasses counterexample

A logging-only capture at e4f8329d records the 62 continuous states immediately before and after accepted-step and off-point manifold projection. The complete three-second output trace is byte-identical to the earlier diagnostic trace (c64dc8502b20f6d0d263e1e7ab614ebd63895b3bc7c61887d5616dd62a83eedb). The instrumented solver budget was 30 seconds to accommodate logging; numerical tolerances were unchanged. This supplies neither ordinary timeout nor coverage or performance credit. All temporary production logging has been removed.

Of 3,137 accepted-step projection calls, 622 retain a state correction. The first retained correction, at time 0.15539131879412615, increases the difference between the two plants’ angular velocities from 2.40e-8 to 2.40e-7. At time 1.2223001460204923, that difference increases from 1.40e-6 to 5.25e-6 in one projection. The rectangular projection blocks include the independent revolute angles and velocities as movable coordinates: 23 constraints act on 25 states in one plant and 29 constraints on 31 states in the other. The runtime’s minimum-norm correction therefore moves those independent coordinates differently for the two representations.

Off-point observation projection alone does not explain the force counterexample. At time 1.986, the angular-velocity disagreement is already -8.863e-6 before observation projection and becomes -8.865e-6 afterward. Combined with the same-state OMC force agreement recorded below, this establishes accepted-step projection as a contributor to the trajectory error. It does not yet establish that projection is the sole contributor, or certify a repair.

All 26 remaining state declarations were states in the original DAE; none is a surviving algebraic promotion. A small experiment established that final state selection can miss a retained value definition for an artificial lifted rate, but that narrower case does not explain this model’s remaining states. Its failing/passing regression and trial patch are archived, and the trial production change was removed. The next investigation must follow original state constraints through differentiation and final state selection, preserving their equations and initialization obligations. Changing observation sampling alone or adjusting tolerances cannot close the four force-channel deviations.

Evidence is under rolling-wheel/line-force-rate-projection-*, rate-projection-*, and retained-state-selection-*. No new cohort claim, baseline promotion, or counterexample closure is made. Combined quick/full verification and complete MultiBody support remain outstanding.

Independent explicit rates close structural reduction; force differences remain

Structural value reconstruction now follows a unique explicit derivative definition when its right-hand side closes acyclically onto source states and invariants. Each definition carries its residual owner and expression together; holonomic preflight excludes the equation being replaced. The value proof and builder replay that same definition. This extends SPEC_0007’s structural substitution profile and its STRUCT-T03 catalog under MLS Appendix B and section 8.3.1, while retaining whole tensor owners and original equations.

The regression der(x)=v; der(v)=-x; q=2*der(x) now retains q=2*v in the value manifold, without derivative reads. Sizes 1, 3, and 4,096 have the same IR size. Controls refuse missing, cyclic, and duplicate rate definitions and prevent either dynamics equation from replacing itself. All 195 structural tests and structural Clippy pass. The duplicate-definition fixture was corrected to construct the detector’s explicit subtraction form; Clippy required moving a nested test choice outside its enclosing conditional. OMC’s backend XML independently produces q[i]=2*v[i] for all three fixture coordinates.

LineForceWithTwoMasses now reaches a structurally prepared DAE with 22 retained manifold expressions, 697 continuous owners, and 26 state declarations. The artifact-enabled worker passes initialization and simulates through the complete three-second interval. Its four force-difference channels prevent a parity claim: the canonical comparator checks 879 channels, with 875 high and four severe (body_f_diff[1:2], rod_f_diff[1:2]). All initialization channels compare high. The maximum observed force difference is about 1.015e-4, while OMC’s original trajectories keep the same differences near roundoff. This is an active refinement counterexample, with no exception or coverage credit.

The next comparison separates force calculation from trajectory error. At sample times 1.986 and 2.028, OMC is initialized with the four independent revolute states from Rumoca’s saved trace, preserving the original numerical tolerance. OMC then reproduces Rumoca’s accelerations within 2.1e-14 and forces within 3.1e-13. Rumoca’s two equivalent plants have angle differences up to 2.52e-6 and velocity differences up to 1.51e-5. Their retained position relations hold to roundoff and sampled velocity relations to about 1.8e-8. These checks point to integration or state projection rather than incorrect instantaneous force equations. Rumoca integrates 62 scalar states versus OMC’s four; the next capture will distinguish native trajectory drift from corrections made during accepted-step and off-point projection.

Tier 1 multibody-rate-values-{frontier,origin,canary} retains all prior bands and phase/execution statuses: four, six, and nine high comparisons, with one, two, and zero reviewed exclusions, and no missing or nonidentifiable traces. The ordinary LineForceWithTwoMasses attempt still exceeds the unchanged Solve budget, so the diagnostic completion is not cohort coverage. No new complete cohort claim is made. The active force counterexample blocks unrelated capability work and release; combined quick/full gates and the full MultiBody goal remain outstanding. Evidence lives under rolling-wheel/rate-values-*, line-force-rate-values-comparison-1, and line-force-omc-same-state-1.

Complete comparison after source-equation exclusion repair

The complete multibody-lift-owners-full-11 run at 6da1462ae77aa9d281bd2e384822d94a69644f15 passes with 165/566 strict-high models (29.15%), 165 compared, 20 reviewed exclusions, and zero missing, nonidentifiable, near, or deviating comparisons. All 22,976 initialization channels are high. Raw execution remains 185; MultiBody remains 23/42 high with one reviewed BevelGear1D exclusion. Every previous comparison band and completed execution is retained. One already-failing Media model changes its reported failure phase from Flatten to ToDae; this earlier-phase diagnostic change earns no execution or coverage credit and is not attributed to the structural repair.

The run uses 11 workers, fresh Rumoca executions, and cached unchanged OMC references. Tracked source is clean at the measured commit; the dirty flag comes from the foreign untracked comm_fastdyn.md. The full previous/baseline receipts are rolling-wheel/lift-owners-full-{previous,baseline}-delta-1.json. No baseline is promoted. Combined verify quick and verify full remain outstanding.

A subsequent saved-source probe narrows the next LineForceWithTwoMasses gap. der(jointUPS.rAxis_0) has a unique explicit definition, body2.v_0-body1.v_0, whose value proof succeeds with those two state anchors. The value proof nevertheless rejects the derivative coordinate itself. That blocks the independent der_rAxis_a_L and angular-velocity definitions, leaving the circular auxiliary reconstruction that the committed repair now excludes. OMC’s corresponding equations use the translational velocity directly. rolling-wheel/line-force-lift-owners-rate-facts-1.json records the exact source identities. The temporary diagnostic test passed and was removed. The next repair must replay independent explicit rate definitions while retaining their source owners and rejecting use of a dynamics equation to replace itself.

Algebraic lifts exclude their source equation from auxiliary reconstruction

The exact pre-lift capture for LineForceWithTwoMasses identifies a source-owner mix-up in the structural proof walk. Owner 434 has residual expression 6022 and right-hand side 6021. Its auxiliary inverse for jointUPS.R_ia_a.w records 6022, but algebraic-lift preflight excluded 6021. It could therefore reconstruct through the equation it was replacing, losing the independent angular-velocity relation. The differentiated expression and excluded equation now have separate roles: preflight excludes the source residual, while the derivative proof stays bound to the right-hand side. Ordinary holonomic constraints retain their existing exclusion. This implements SPEC_0007’s structural-lowering ownership contract and STRUCT-T03 under MLS Appendix B and section 8.3.1.

A small valid source model reproduces the defect with u=der(x), v=shift(u,zeros(n)), q=forward(v), der(q)=a, and a=-q, where the pure functions add the supplied offset and forward their argument. Its rejected lift would replace the relation between der(x) and q with a circular alias. The regression proves that its auxiliary block owns the source equation and must not supply that equation’s lift. An independent u=x, der(x)=ones(n) control still lifts and retains the relation between v and x. Both tests pass for sizes 1, 3, and 4,096; the positive reconstruction has constant IR size. All 192 structural tests and all-target/all-feature structural Clippy pass. The initial Clippy pass required extracting a nested test traversal into a helper. OMC accepts both small models and preserves their expected differential equations; its backend XML is retained under rolling-wheel/lift-owners-omc-1.

The rebuilt original model has byte-identical source DAE and retains owner 434’s undifferentiated Frames.absoluteRotation relation. It still fails Solve with EL005: the deepest snapshot matches 2,201/2,204 rows. The remaining rows 2171–2173 now map to owner 688, rod1.frame_b.R.w=jointUPS.frame_a.R.w, with unmatched unknowns jointUPS.f_bd_a[1:3]. This is diagnostic evidence, not a simulation or coverage gain. Next, trace the independent w_rel_ia1 definition through der_rAxis_a_L and the explicit derivative of rAxis_0, comparing each substitution against OMC’s translational-velocity equations.

Tier 1 multibody-lift-owners-{frontier,origin,canary} retains all prior bands and phase/execution statuses from the demotion-owner controls. The fixed target sets compare four, six, and nine high traces respectively, with one, two, and zero reviewed exclusions. There are no missing, nonidentifiable, near, or deviating comparisons. The ordinary LineForceWithTwoMasses attempt still exceeds the unchanged Solve budget. rolling-wheel/lift-owners-evidence-1.json binds the source captures, repair, tests, OMC equations, and gate deltas. Combined quick/full verification and the full MultiBody goal remain outstanding.

Complete cohort retains all prior results after equation-activity checks

The full multibody-demotion-owners-full-11 comparison at dd3143d1ad006b9ae020a0dbcc3067e8b2d71f13 passes: 165/566 strict-high (29.15%), 165 compared, 20 reviewed exclusions, zero missing/nonidentifiable traces, and zero near/deviating comparisons. All 22,976 initialization channels are high. Raw execution remains 185 models; the exclusions receive no strict-high credit. MultiBody remains 23/42 high, with one reviewed BevelGear1D exclusion. Every model retains its preceding band and phase/execution status, including the previously repaired counterexamples. The ordinary LineForceWithTwoMasses attempt still exceeds the unchanged 10-second Solve budget.

The run uses 11 workers and fresh Rumoca execution; unchanged OMC references are reused by the comparator. Tracked source is clean, and comm_fastdyn.md remains foreign and untracked. Receipts rolling-wheel/demotion-owners-full-{previous,baseline}-delta-1.json bind the complete identical rosters and comparator artifacts. The aggregate gate exits zero. No baseline is promoted. Combined verify quick and verify full remain outstanding, and complete MultiBody support is not established.

Structural candidate acceptance rejects reflexive coordinate equations

The next capture isolates the remaining circular substitution at holonomic round 20, residue 24. Demoting body1.w_a to jointUPS.frame_ia.R.w differentiates the latter through its existing definition body1.z_a. Both states already have that same derivative definition, so the original z_a=der(w_a) equation becomes reflexive. The source derivative-cycle walk need not re-enter the demoted coordinate: its algebraic chain reaches body2’s derivative instead.

A small source model, x=y; der(x)=a; der(y)=a, reproduces the false structural success. OMC rejects it as structurally singular at a[i]=a[i]; a companion model with independent derivative coordinates and a=-y compiles to the expected first-order motion equation. Rumoca’s direct and holonomic candidates now share an acceptance check that refuses newly reflexive equalities of Real coordinates. It compares branded coordinate identity, including distinct source expression occurrences, and retains source equations and tensor domains. This does not rewrite numerical residuals, eliminate function calls, or sample rank.

Both regression models pass for scalar, singleton, three-element, and 4,096-element payloads. All 190 structural tests and all-target/all-feature structural Clippy pass. The first Clippy run identified a function two lines over the size limit; the direct reconstruction checks were factored into a helper, then tests and Clippy reran successfully. Temporary capture tests were removed and archived outside production source.

LineForceWithTwoMasses now refuses Solve construction with EL005 instead of entering initialization with the known tautology. Its deepest saved reduction matches 2,201/2,204 equations; the three unmatched rows are owner 545’s original angular-velocity relation. The unmatched unknowns are jointUPS.f_c_a[3] and jointUPS.f_bd_a[2:3]. The returned error still describes the initial 2,164/2,204 matching; the diagnostic snapshot preserves the later obstruction. The next comparison follows JointUPS’s explicit w_rel_ia1 construction from its translational kinematics to identify the missing independent reconstruction. No simulation or parity gain is claimed.

Tier 1 multibody-demotion-owners-{frontier,origin,canary} retains every band and execution status from the corresponding function-zero controls. The six frontier targets compare four high traces with one reviewed exclusion; the eight origin targets compare six with two reviewed exclusions; the fixed 20-model canary compares nine with none. Every comparison is high, with zero missing, nonidentifiable, or deviating traces. These partial runs are regression evidence. rolling-wheel/demotion-owners-evidence-1.json binds the captures, OMC probe, source repair, tests, inspector, and gate deltas. The complete cohort below remains authoritative; combined quick/full verification is still outstanding.

Exact identity and zero derivatives expose the next LineForce failure

Two structural producer repairs now preserve exact identities through tensor reconstruction. A construction-proved identity coefficient in A*q=b emits the RHS directly at each admitted derivative order. Other coefficients retain the checked aggregate solve, including tunable parameters whose default is one. Supplied function derivatives now respect the existing checked zero-expression proof in both admission and construction. A zero-default parameter remains a parameter; its numeric default does not establish an identically zero function.

The source-backed regressions fail before the respective repairs and pass after them for vector extents 1, 3, and 4096, with constant IR size. All 188 structural tests, all-target/all-feature structural Clippy, formatting, and whitespace checks pass. The ordinary workers rebuild successfully. The OMC ZeroRotation probe, using MSL Frames.resolve2(R, zeros(3)), retains q[i]=y[i] and der(y[i])=-y[i], with no spurious rotation derivative in those equations.

The actual LineForceWithTwoMasses source DAE remains byte-identical to the original capture. The identity-only repair removes its synthetic identity solve but leaves the circular angular-velocity derivative. With the exact-zero repair, prepared owner 434 becomes der(jointUPS.frame_ia.R.w)-body1.z_a, matching the previously captured OMC relation. Owner 545 retains the original angular-velocity value equation and demotes body1.w_a; it is no longer replaced with the circular acceleration residual. The diagnostic has 31 state declarations and 27 manifold rows. These counts are not OMC’s scalar-state count.

The model still fails initialization with EX002 and produces no trace. The next capture identifies local block 60, logical row 1635, and body1.z_a[1] (Y 1399). All 166 failure events share one numerical point and a zero 1-by-1 matrix. Source owner 546 is body1.z_a-der(body1.w_a); its prepared residual is exactly body1.z_a-body1.z_a. Solve program 522 loads the vector once and subtracts the same registers from themselves, producing logical rows 1635–1637. This is a remaining circular substitution in state demotion, already present before Solve; it is not evidence of an AD-only defect. The next proof must trace the derivative anchor that permits this self-substitution. Solve still takes 12.623 seconds in the artifact-enabled diagnostic, exceeding the ordinary 10-second budget.

Tier 1 runs multibody-function-zero-{frontier,origin,canary} retain every band and execution status from the corresponding demotion-bounds controls. The six frontier targets compare four high traces with one reviewed exclusion; the eight origin targets compare six high traces with two reviewed exclusions. The fixed 20-model canary compares nine high traces, with no exclusions. All three have zero missing/nonidentifiable/near/deviating comparisons. These partial runs are regression checks, not new cohort coverage. Receipts rolling-wheel/function-zero-{evidence,next-failure}-1.json bind the source, generated equations, OMC probe, tests, worker captures, and comparator deltas. The full cohort below remains authoritative: 23/42 MultiBody high. Combined verify quick and verify full remain outstanding, and no PR is opened.

LineForce sensitivity failure is a circular structural equation

The failure-only debug event identifies the exact singular block in LineForceWithTwoMasses. All 172 failure events share one Y/P/time point and one zero 3-by-3 matrix. The selected logical rows are 1634, 1632, 1633 and the unknowns are body1.z_a[2], body1.z_a[1], body1.z_a[3]. Replaying the saved Solve model identifies global block 886, with a full structural pattern. The local refresh block number is 186; those two inventories differ.

The rows belong to the differentiated w_a = Frames.angularVelocity2(frame_a.R) equation, not the torque equation. The initial zero-inertia torque hypothesis is rejected by this exact row mapping. Source DAE owner 545 contains body1.w_a - angularVelocity2(body1.frame_a.R). The same owner in the structural DAE contains:

body1.z_a - (
  Internal.resolve2_der(jointUPS.R_ia_a.T, jointUPS.R_ia_a.w,
                       jointUPS.frame_a.R.w, zeros(3))
  + linear_solve(identity(3), body1.z_a))

The acceleration cancels through the generated identity-matrix solve. At the captured point, independent unit-coordinate perturbations of the primal residual and the projection, full, and directly regenerated JVPs all return zero on these rows. This rules out missing sparse entries or an AD-only error as the cause of the observed zero matrix. The circular form already exists before Solve lowering; the next proof must isolate which source-constraint/auxiliary reconstruction introduces it and reproduce that transformation in a focused regression.

The saved OMC backend instead retains $DER.jointUPS.R_ia_a.w[i] = body1.z_a[i] (equations 310–312) and the equivalent body2 derivative relation (320–322). Its four selected scalar states are the two revolute angles and their angular velocities. Different state counts alone do not establish the defect; the circular structural equation is the concrete frontier for the next comparison.

rolling-wheel/line-force-circular-sensitivity-evidence-1.json binds the source/structural/Solve artifacts, OMC equations, numerical capture, and public replay. The source DAE is byte-identical to the preceding diagnostic. The new runtime event reports the existing failed matrix without changing evaluation, factorization, seeds, tolerances, or failure classification. All 24 seed-focused tests, all-target/all-feature solver Clippy, formatting, and whitespace checks pass. The worker build and diagnostic helpers complete; the first helper compile needed an external-table slice type correction. These are diagnostic results, not an execution or parity gain. The complete cohort below remains authoritative; combined verify quick and verify full are still outstanding.

Full comparison restores GyroscopicEffects and retains every high model

The complete multibody-demotion-bounds-full-11 run at b008b1a794d951e17e1d9b2e9fc7ccc9a335e387 compares 165/566 strict-high (29.15%): 165 compared, 20 reviewed exclusions, zero missing/nonidentifiable traces, and zero near/deviating comparisons. All 22,976 initialization channels are high. Raw execution is 185 models; exclusions are not high-parity credit. The run uses 11 workers and the unchanged phase/solver budgets. Its worktree digest is 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35; tracked source is clean and the foreign comm_fastdyn.md remains untracked.

MultiBody returns to 23/42 high, with one reviewed BevelGear1D exclusion and no missing comparisons. GyroscopicEffects is restored from the preceding full run’s Solve timeout to high parity. Every previously high model and reviewed excluded execution is retained. The four repaired counterexamples stay high; both Cauer examples retain their gains against the preceding committed baseline. RevoluteConstraint advances from a timeout to an explicit structural singularity, which is still a failed model. LineForce still times out in Solve.

Receipts rolling-wheel/demotion-bounds-full-{previous,baseline}-delta-1.json compare identical 566-model rosters against the failed independent-lift attempt and the preceding target/msl/results baseline. The aggregate gate exits zero. No baseline is promoted. This closes the lost-model regression and permits returning to the LineForce frontier; 100% MultiBody and combined quick/full verification remain outstanding.

The current LineForce diagnostic confirms two separate failures. Its exact saved source prepares in 10.348 seconds with 32 state declarations, 28 manifold rows, and 954 candidate reconstruction attempts. The artifact-disabled actual worker spends 12.870 seconds in Solve, then fails initialization with EX002: algebraic projection sensitivity matrix is singular. It produces no trace. The diagnostic’s larger outer budget grants no coverage credit; the ordinary full run still fails the unchanged 10-second Solve budget.

The producer artifacts, worker hash, and perf capture are bound by rolling-wheel/line-force-demotion-bounds-evidence-1.json. Perf attributes cost to repeated incidence projection, variable reservation/equality, allocation, and hashing; it does not establish that optimizing name lookup alone would meet the budget. The next correctness investigation must identify the singular projection block and its exact initialization point, then compare those equations with the saved OMC backend equations. OMC explicitly retains the line-length first/second derivative chain and damper force closure (line-force-omc-damper-closure-1.json). The current evidence does not yet prove which Rumoca equation or coordinate diverges from that closure. No further compiler/runtime change is made on this hypothesis alone.

GyroscopicEffects matching bound passes focused verification

The first divergent search decision precedes the independent additive-lift fix: the necessary independent demotion-anchor repair expands first-round admission from 27 to 57. Earlier affine/value-identity inspections emit 364 records versus 528 after that repair. The same-source prepared systems before/after the final additive-lift change remain identical. Artifact-disabled saved/current workers measure Solve at 5.165/6.952 seconds in one diagnostic pair; the saved worker is the affine-state-anchor build, not the profiling script’s current Git HEAD. Hashes, selected records, and timings are retained in rolling-wheel/gyro-demotion-search-localization-1.json.

Direct search already retains the first checked strictly reduced candidate. Later candidates can supersede it only by completing the matching. The candidate optimization bounds how many continuous scalar rows a demotion can change, counting whole compact owners and conservatively adding any restored lifted owner. Changing at most m rows can increase maximum matching cardinality by at most m: removing those rows from any new matching leaves a source matching. Direct demotion preserves equation/unknown counts, so residue cannot decrease by more than 2*m. Candidates with a larger source residue need no trial once a checked reduction exists. No candidates are removed from admission, no state priorities change, and every attempted/selected result retains its checks.

The source-backed five-independent-constraint regression fails on the original search with 15 reconstructions and passes with the bound for scalar, singleton, and three-element tensor payloads. A separate mixed-width fixture proves that a later Sorted tensor candidate still supersedes a Reduced scalar candidate, producing the exact same prepared DAE as its exhaustive trial. All 184 structural tests and all-target/all-feature Clippy pass (demotion-bounds-green-5.log, demotion-bounds-clippy-3.log). Earlier fixture naming and nesting failures remain recorded. No debug probe, model branch, timeout change, or tolerance change is used.

GyroscopicEffects retains byte-identical prepared DAE and manifold artifacts and all 24 selected reductions. Actual reconstruction attempts fall from 452 to 281; total observation records fall from 528 to 357. Its public structural diagnostic falls from 5.287 to 3.097 seconds. A sequential artifact-disabled actual-worker pair measures Solve at 6.865/5.043 seconds and integration at 0.407/0.405 seconds. Both execute and initialize successfully. These are single diagnostic pairs, not statistical benchmark claims. Perf capture completes with 5,062 candidate samples; its deliberate SIGINT exit is recorded separately from the worker’s successful exit. Receipts are gyro-demotion-bounds-{artifact-delta,ordinary-pair}-1.json.

At a671ecb8, digest 4bdbfb623dc4082fb3a43e9c4ef67810c2733a35b40b48ae97fa7014a578a2a1, the ordinary multibody-demotion-bounds-frontier run has four high comparisons, one reviewed BevelGear1D exclusion, and zero missing/deviating results. It restores GyroscopicEffects execution relative to the latest full attempt and preserves every phase/band against the same six-model multibody-round-facts-origin roster. LineForce still times out in Solve. The eight-model origin run preserves all six high comparisons and both reviewed exclusions. The fixed canary preserves all twenty phase/band rows, nine high comparisons, zero exclusions/missing/deviation. Their deltas are demotion-bounds-{frontier,origin,canary}-delta-1.json.

A fresh complete cohort is still required to restore coverage credit; the latest authoritative MultiBody count below remains 22/42. Combined quick/full verification also remains outstanding; this checkpoint is not release evidence.

Full comparison closes false traces but loses GyroscopicEffects execution

The full multibody-independent-lift-full-11 run at a671ecb8ac35bfc10a3a9db02e0f81feb40f651c compares 164/566 strict-high (28.98%): 164 compared, 20 reviewed exclusions, zero missing/nonidentifiable traces, and zero near/deviating comparisons. All 22,009 initialization channels are high. PreLoad and DCPM_Drive retain excluded execution; the four repaired counterexamples remain high. Both Cauer models gain high comparison against the preceding committed cohort. The run’s worktree digest is 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35; the only untracked workspace file is the preserved comm_fastdyn.md.

GyroscopicEffects loses its high result to a Solve timeout: the worker exceeds the unchanged 10-second budget after 11.559 seconds. MultiBody consequently falls to 22/42 high: 22 compared, one reviewed BevelGear1D exclusion, and no missing comparisons. The aggregate gate exits successfully, but the loss of a previously high model blocks campaign acceptance and baseline promotion. The complete phase and band delta is rolling-wheel/independent-lift-full-delta-1.json. LineForce remains a Solve timeout and earns no new credit.

The first GyroscopicEffects diagnostic rejects an enlarged structural system as the explanation for the independent-lift change. Public inspections of the identical source DAE produce byte-identical prepared DAEs and manifold ordinals before and after that change, with all 528 reduction records identical. Their diagnostic reduction durations are 5.324 and 5.287 seconds; this single pair is not a speed claim. The prepared DAE SHA-256 is 38126fc10c1eb1c89c6b0cad70feed5efb3c1d4ef33396dad93d8216ac3a6dad. The receipt is rolling-wheel/gyro-independent-lift-artifact-delta-1.json.

An artifact-enabled actual-worker run completes simulation and records 7.545 seconds for Solve lowering. Its perf interval also includes artifact serialization, so that profile cannot be treated as ordinary cohort timing. The profile exposes existing variable-name equality and reservation costs, incidence projection, allocation, hashing, and output writes; it does not establish additive-proof search as the bottleneck. Artifact-disabled control and candidate worker profiles are the next comparison before choosing a performance change. The failed full attempt remains recorded; a diagnostic completion does not restore its coverage credit.

Independent source equations restore PreLoad execution

The new additive lift proof excludes the equation being replaced and the coordinate being lifted. It derives the value from other source equations, refuses cycles, and reconstructs the value and derivative from that same postordered derivation. Rows are collected once per immutable source round; payloads remain whole tensors. Non-additive lift profiles are unchanged.

The source-backed regression demonstrates the previous semantic loss directly: r+p=0; s=L-q; p*q=-1; r=q admitted a lifted manifold at p=2, q=-0.5, r=-2, s=1.5, L=1, although the original r=q residual is -1.5. The repaired manifold rejects that point. Scalar, one-element, and three-element payloads, independent signed aliases, circular aliases without an independent definition, and restoration after demotion are covered. All 182 structural tests and all-target/all-feature Clippy pass. Earlier fixture, shape, and lint failures remain in the numbered logs; final evidence is independent-lift-green-5.log and independent-lift-clippy-3.log.

The ordinary multibody-independent-lift-origin run at 78b2018f, worktree digest 80994095174927ccb78e0af45be059e0e8645b32f083db56336ba46e1216e4d0, restores PreLoad from EL005 to sim_ok/ic_ok under its existing reviewed comparison exclusion. DCPM_Drive retains its reviewed exclusion. All six compared traces remain high: Damper, ElastoGap, Oscillator, ArmatureStroke, CauerLowPassAnalog, and CauerLowPassOPV. There are zero missing, nonidentifiable, near, or deviating comparisons. This is focused evidence; the repaired counterexamples remain required in the next complete sweep.

The exact saved PreLoad source now prepares with eight states and two manifold rows. The transformed velocity equation is friction.v_rel - spool.v, matching OMC’s generated spool.v = friction.v_rel. No initial-value pins are added. The prepared DAE has SHA-256 dcf1430b5b2b91ab2650f4c8f7b0708922b4d6af85e1e003cb2cf0828d52ebfb. The public inspector is bound to the ordinary worker’s exact library artifacts; its timing is diagnostic only. Receipts are preload-independent-lift-equations-1.json and independent-lift-inspector-receipt-1.json under rolling-wheel/.

The fixed multibody-independent-lift-canary preserves every phase and band against multibody-lift-restoration-canary: nine compared high, zero exclusions/missing/nonidentifiable results, and all 175 initialization channels high. The other eleven models remain unsuccessful. Origin and canary deltas are in independent-lift-{origin,canary}-delta-1.json. This closes the focused execution regression; full-cohort and combined quick/full verification remain outstanding, and no new MultiBody coverage or RollingWheel speed claim is made.

Focused counterexamples recover; PreLoad still blocks the candidate

The latest multibody-lift-restoration-origin run at 78b2018f, with worktree digest dc6a6448d181197d9da861754ded24f6fa8e429ab85e4717dde6a4a96911cccb, compares all six available traces high: Damper, ElastoGap, Oscillator, ArmatureStroke, CauerLowPassAnalog, and CauerLowPassOPV. DCPM_Drive executes under its existing reviewed comparison exclusion. There are zero missing, nonidentifiable, near, or deviating comparisons. PreLoad still fails structural reduction with EL005, reporting 111/116 matches. Its prior excluded execution must be restored before this candidate is complete. These focused results do not establish a new full-cohort result or close the required full-sweep check.

The repairs separate exact value identities from displaced derivative identities and preserve the proof used by reconstruction. Direct demotion selects an independent derivative anchor. Retained first derivatives and lifted definitions replay their captured exact anchor policy, including in cache keys. Armature’s two canceled velocity conditions now retain actuator/load velocity differences, matching the relationship in OMC’s generated equations. Finally, undoing a lift restores the original source residual instead of a substituted manifold expression. PreLoad’s source connection at owner 108 is now restored as rod3.flange_b.s - rod4.flange_a.s; it no longer duplicates owner 109, rod4.flange_a.s - spool.flange_a.s.

All 180 structural tests and all-target/all-feature Clippy pass. The fixed multibody-lift-restoration-canary preserves every phase and band against multibody-affine-state-anchor-canary: nine compared high, zero exclusions, missing, or nonidentifiable results, and all 175 initial-condition channels high. Eleven other canary models remain unsuccessful; this is preservation, not complete canary model support. Receipts under rolling-wheel/ are lift-restoration-{origin,canary}-delta-1.json, lift-restoration-green-1.log, lift-restoration-clippy-build-1.log, lift-anchor-manifold-delta-1.json, and preload-lift-restoration-equation-delta-1.json.

The remaining PreLoad hypothesis is an earlier circular dependency: a lifted definition’s exact value anchor may rely on the equation being replaced. Its derivative then duplicates the neighboring connection, and the next holonomic replacement correctly refuses a vacuous residual. Restoring the original equations fixes the subsequent ownership loss but does not establish an independent lifting proof. The next reproduction must distinguish a value definition supported by other equations from one supported by its own owner. No tolerance, model-specific branch, or runtime dispatch change was made. Changes remain uncommitted; a new full sweep and combined quick/full verification are outstanding while this execution regression is unresolved.

Uncommitted state-balance work exposes circular manifold reconstruction

The state/invariant balance regression now passes: a fixed support contributes no derivative, while only a proved zero support may disappear from an exact value equality. LineForce’s identical source DAE now reaches a prepared system instead of stopping at 2,203/2,204 matches. Its ordinary worker still exceeds the 10-second Solve budget. An unrestricted diagnostic worker on the first candidate reaches initialization and reports a singular algebraic sensitivity matrix; it produces no trace and earns no coverage credit.

The first full candidate, multibody-state-balance-full-11 at 78b2018f with worktree digest 50b444de5b5b0062fbc72ac2279a70e0a521db2422ebf62c1978a0bae5e06f1f, compares 165/566 high, with 19 reviewed exclusions and zero missing or deviating traces. CauerLowPassAnalog and CauerLowPassOPV gain high comparison, but PreLoad loses its previously executable, excluded result. This execution loss remains a regression despite the gate passing.

Changing derivative lookup to the full affine class restores PreLoad’s structural replay, but the next full gate fails. At the same HEAD with digest 842e1538ad40459344defcf99cf448f59f21d3851ccdb0d5912cac56e33594d9, multibody-affine-state-anchor-full-11 compares 165 traces: 161 high, one near, three deviating, 19 reviewed exclusions, and zero missing. Damper, ElastoGap, Oscillator, and ArmatureStroke are actionable counterexamples. DCPM_Drive also loses its previously executable, excluded result. Both focused six-model runs and the fixed canary had unchanged bands; all 175 structural tests and affected Clippy passed. The full evidence overrides those focused results.

The exact Damper replay localizes the first loss: retained manifold expressions 390, 400, and 410 become massN.flange_b.s - massN.flange_b.s. The other three manifold rows retain the body half-length offsets. At initialization the actual worker reports mass1.flange_b.s=3.5 and damper1.flange_a.s=4.5, violating their source connection by one metre. A source equality used to materialize its own replacement has erased the independent position constraint. This requires a structural proof repair; numerical tolerances cannot repair it. The candidate remains uncommitted and further coverage work is stopped.

Receipts under rolling-wheel/ are state-balance-full-delta.json, affine-state-anchor-full-delta.json, preload-{before,after}-inspection-1.log, and affine-state-anchor-inspector-receipt-1.json. Damper’s exact artifacts and prepared manifold are in damper-affine-state-anchor-{artifacts,inspection}-1/. No solver dispatch change was made. Affine value recovery already uses guarded tearing after full AD assembly; SeedBlockLinearization still factors the full matrix for sensitivities. A compiler-composed reduced value/JVP kernel is pending.

Retain stalled equations for the LineForce comparison

Structural inspection now retains one already-owned stalled DAE together with its manifold and exact failure at the point it would otherwise be discarded. The ordinary observer drops the same values without allocating, cloning IR, or repeating analysis. This diagnostic snapshot is never a prepared model; successful preparation returns no stalled snapshot.

The LineForceWithTwoMasses capture reaches 2,203 matched equations out of 2,204 equations and unknowns, with 26 retained manifold expressions. Its seven-equation overdetermined witness ends at f_x[1475], the connection jointUPS.axis.s = damper1.flange_b.s; jointUPS.f_bd_a[2] remains unmatched. All 1,176 recorded reduction events are byte-identical to the prior public inspection, showing that observation preserved the actual decisions. The snapshot has SHA-256 f59b86731aa723b8db8523cc02f86012ebbf131c4a4ceccb754c9e5cc331be29. Receipts are rolling-wheel/line-force-stalled-capture-1.json and rolling-wheel/stalled-diagnostic-helper-receipt-1.json; the captured DAE and manifold are under rolling-wheel/line-force-stalled-inspection-1/.

All 174 structural tests and all-target/all-feature Clippy pass. The new regression compares observed and ordinary failures, preserves the source DAE, and checks the retained failure against a fresh analysis of that exact DAE. The first build found an existing recorder test needing the changed private signature; Clippy then required extracting the discard operation to keep the entry point within its function budget. Final evidence is stalled-diagnostic-clippy-2.log and stalled-diagnostic-libraries-2.jsonl. This is diagnostic progress, not a new model pass or a worker timing result.

Quick comparison preserves the cohort; spec budget repaired

The full MSL step of verify quick, written to target/msl/results at 648ec46e0806afcdbdbcc5ef7c7b81e2485d5f70, preserves every comparison band: 163/566 strict-high, 163 compared, 20 reviewed exclusions, zero missing or nonidentifiable traces, and zero near/deviating comparisons. All 22,718 initialization channels are high. The only phase/status change from the preceding named full sweep is Fourbar_analytic moving from timeout to explicit structural EL005; it gains no parity credit. The receipt is rolling-wheel/round-facts-quick-msl-delta.json.

Workspace lint, all 28 corpus-pin cases, and all 243 architecture tests pass. The adjacent spec gate fails because SPEC_0036 and SPEC_0038 exceed the 2,500-word cap. Redundant prose is condensed with requirements and normative catalog links preserved, bringing them to 2,498 and 2,495 words. The existing gate executable’s exact test_specs_respect_size_budgets test now passes (rolling-wheel/spec-budget-repair-focused-1.log). This documentation repair was applied after the architecture step while the unchanged compiler was building workspace tests. The workspace step subsequently passed all 7,747 tests with zero skipped, plus its documentation tests. A complete rerun of the repaired architecture command passed all 243 architecture tests and 17 suite gates (spec-budget-architecture-recheck-1.log). The original combined quick run records its one failed step; combined full remains outstanding.

Full sweep restores Driving without losing a high model

The complete multibody-round-facts-full-11 sweep at 720680e9ed3055308b2516271bbbb5d96a200cc1 passes its quality gate: 163/566 strict-high models (28.80%), 163 compared, 20 reviewed exclusions, zero missing or nonidentifiable traces, and zero near/deviating comparisons. All 22,718 compared initialization channels are high.

MultiBody reaches 23/42 high: 23 compared, one reviewed BevelGear1D exclusion, and zero missing comparisons. RollingWheelSetDriving returns to high with Solve completing in 6.1343 seconds under the unchanged 10-second budget. No previously high model loses its band. The only other phase/status change against multibody-alternative-definitions-full-11 is DC_CompareCharacteristics moving from timeout to an explicit structural refusal; it gains no compilation or parity credit. LineForceWithTwoMasses still times out in Solve. The complete delta and artifact digests are in rolling-wheel/round-facts-full-delta.json.

The Driving cohort regression is closed by this complete comparison. The next compiler frontier remains LineForceWithTwoMasses’ unresolved displacement and derivative closure against the retained OMC equations. Combined verify quick/full remain outstanding; this successful MSL sweep does not establish those gates, complete MultiBody support, or PR readiness.

Share differentiation facts within each immutable reduction round

RollingWheelSetDriving’s saved pre-change and current workers produce identical Flat, DAE, structural DAE, Solve IR, and complete trace bytes. The timeout did not arise from a larger generated system. A public reduction inspection spends 6.546 seconds selecting 24 transformations from 418 attempts. Actual-worker perf samples locate repeated source analysis inside preparation. These records are in rolling-wheel/wheel-set-round-facts-diagnosis-1.json.

Each reconstruction previously collected the same source differentiation facts again. A private ReductionSource now binds one immutable DAE to its facts; candidate discovery, manifold preflight, and all attempted reconstructions in that round consume that pair. The next transformed source constructs fresh facts. Candidate order, matching, retained equations, initialization checks, and numerical execution are unchanged. SPEC_0007 records the ownership rule; its wording was also condensed to satisfy the 2,500-word spec budget without removing the governing STRUCT-T03 catalog requirements.

The two-independent-constraints regression records three collections instead of one before the fix, then one per round and a fresh collection after the source changes. All 173 structural tests and all-target/all-feature Clippy pass (round-facts-red-1.log, round-facts-green-2.log, and round-facts-validation-build-2.log). Earlier signature and lint failures remain in the numbered logs. The ordinary worker build includes msl_tests.

One alternating ordinary-worker pair with artifact output disabled measures Solve program preparation at 8.4275 seconds for the saved parent worker and 5.7507 seconds for the candidate. This observed 31.76% reduction is one unpinned pair, not a repeated benchmark claim. Simulation remains about 0.333 seconds for both. The artifact-enabled candidate also retains all five complete artifacts byte-for-byte, including the trace with SHA-256 268a5d61f8adf56e17db4a2661b93593249b4cb3a112a25c0990af7d6826955d. Receipts are round-facts-driving-timing-1.json and round-facts-driving-artifact-delta-1.json.

The timing field named ir_solve_structural_dae_seconds currently receives runtime_value_seconds; it does not measure structural index reduction. ir_solve_lower_seconds includes prepare_for_solve plus program/artifact lowering. The approximately 47 ms field in earlier diagnostics must not be interpreted as the cost of index reduction.

The six-model multibody-round-facts-origin gate compares RollingWheelSetDriving, RollingWheel, GyroscopicEffects, and OvervoltageProtection high: four compared, one existing BevelGear1D exclusion, zero missing/nonidentifiable/deviating comparisons, and all 2,088 initialization channels high. Driving completes Solve in 6.0957 seconds under the unchanged 10-second budget. LineForceWithTwoMasses still times out in Solve and receives no credit. round-facts-origin-evidence-1.json binds these results.

The fixed multibody-round-facts-canary has no phase or band changes against multibody-alternative-definitions-canary: nine compared high, zero skips, missing, exclusions or nonidentifiable results, and all 175 initialization channels high (round-facts-canary-delta.json). This is focused recovery; the next complete sweep must confirm the cohort regression is closed. Combined verify quick/full and complete MultiBody support remain outstanding. This patch does not implement the reduced AD kernel or establish an OMC runtime speed win.

Alternative definitions gain two high models and expose a timeout

The complete multibody-alternative-definitions-full-11 sweep at f0989e3f1e655485820d5aa267c8d1e76564dcd8 passes its MSL quality gate: 162/566 strict-high models (28.62%), 162 compared, 20 reviewed exclusions, zero missing or nonidentifiable traces, and zero near or deviating comparisons. All 21,826 compared initialization channels are high.

GearConstraint and Modelica.Mechanics.Rotational.Examples.First newly compare high. GearConstraint completes Solve in 3.266 seconds; First completes it in 8.28 ms after previously refusing a 52/54 structural match. However, RollingWheelSetDriving loses its high band through a Solve timeout after 11.103 seconds against the unchanged 10-second budget. Its previous full-run Solve time was 8.675 seconds. MultiBody therefore remains 22/42 high: 22 compared, one reviewed exclusion, and zero missing comparisons. Fourbar_analytic’s structural refusal also becomes a timeout, which establishes no resolution of its structural defect. The complete delta and artifact digests are retained in rolling-wheel/alternative-definitions-full-delta.json.

RollingWheelSetDriving is the next regression to address. A subsequent diagnostic profile completes with 8.1105 seconds of Solve lowering and a 0.3282-second simulation, but that output-enabled, unpinned diagnostic does not replace the bounded full-run failure or establish a performance change. Its artifacts and perf samples are in rolling-wheel/wheel-set-definitions-candidate-1. No reduced-Jacobian kernel or runtime tearing change was made in this patch. Combined verify quick/full remain pending, and complete MultiBody support and PR readiness remain unestablished.

Retain alternative defining equations during structural substitution

LineForceWithTwoMasses’ timeout occurs in structural index reduction, before Jacobian construction. The actual-worker Solve profile ends after 15.3 seconds with 2,164/2,204 equations matched. Inspection records 107 rounds and reaches an intermediate residue of one equation/unknown pair before returning the original singularity. OMC’s retained backend XML instead treats jointUPS.axisLength as a dummy state and supplies its first and second derivatives. Its source has both a geometric length definition and a displacement-plus-offset definition. Evidence is rolling-wheel/line-force-structural-diagnosis-1.json and the retained line-force/omc-stages captures.

That investigation exposed a general distinction: executable causal elimination requires a unique definition, but structural substitution may use an exact state/invariant-anchored equality while retaining the other equations as constraints. Structural facts now complete missing definitions through a deterministic acyclic closure over checked whole-coordinate equations. Every source equation remains; executable causal-definition policy is unchanged. SPEC_0007’s STRUCT-T03 profile records this source-preserving substitution.

The reduced scalar/vector regression fails before the fix and passes afterward. A cycle without value anchors remains unmaterializable. Three older tests that assumed displaced definitions were unsupported now check the admitted source constraints; the constant and parameter-offset cases also evaluate both retained position and velocity residuals at multiple parameter values. All 172 structural tests and affected all-target/all-feature Clippy pass (alternative-definitions-validation-build-7.log). Earlier fixture, oracle, and lint failures remain in the numbered logs.

The five-model multibody-alternative-definitions-origin retains high comparisons for RollingWheel, GyroscopicEffects, and OvervoltageProtection: three compared, one reviewed BevelGear1D exclusion, and zero missing or deviating comparisons. LineForceWithTwoMasses still exceeds the unchanged Solve budget. Replaying the same source DAE with the new reducer records 102 rounds and still ends one equation/unknown pair short; the alternative-definition gap is not the complete explanation. No compilation or parity credit is claimed for that model. Receipts are alternative-definitions-origin-evidence-1.json and alternative-definitions-line-force-delta.json under rolling-wheel.

The fixed multibody-alternative-definitions-canary preserves every phase and band from multibody-invariance-sharing-canary: nine compared high, zero skipped/missing/excluded/nonidentifiable traces, and all 175 initialization channels high (alternative-definitions-canary-delta.json). The next complete cohort and combined verify quick/full remain pending.

Full sweep confirms shared-proof construction

The complete multibody-invariance-sharing-full-11 sweep at f821fcce94d964671e7dcc12e23c2da87cddc4f8 passes its MSL quality gate. It measures 161/566 strict-high models (28.45%): 161 compared, 20 reviewed exclusions, zero missing or nonidentifiable traces, and zero near or deviating comparisons. All 21,865 compared initialization channels are high. The run reports a dirty worktree because the foreign untracked comm_fastdyn.md is present; authored source changes were committed before the run.

GyroscopicEffects returns to high with all 967 channels high; Solve takes 8.771 seconds, including 8.712 seconds of lowering, within the unchanged 10-second phase budget. MultiBody is 22/42 high, with 22 models compared, one reviewed BevelGear1D exclusion, and zero missing comparisons. Its raw simulation completion count is 23/42. IMS_Start remains a reviewed reference boundary with no high-parity credit. No previously high model loses its band.

GearConstraint remains unsuccessful: the prior structural failure becomes a Solve timeout after 11.623 seconds against the 10-second budget. That change does not establish that its structural failure is fixed. The retained invariance-sharing-full-delta.json records both phase changes and the comparison-artifact digests against multibody-guarded-affine-full-11. The earlier failed sweep remains recorded below.

This result establishes neither complete MultiBody support nor a RollingWheel speed win. The affine path uses the checked tearing layout for its reduced linear solve but still evaluates the full block Jacobian first. Its coordinate recovery and original-residual refinement are necessary for the electrical regressions; full-block differentiation is an implementation limitation, not a mathematical requirement of those checks. Combined verify quick/full remain pending; only the full MSL parity gate was run in this sweep.

Share Jacobian invariance proofs across projection blocks

GyroscopicEffects exposed repeated construction work: every algebraic block re-derived operation-invariance facts for the entire canonical Jacobian source. A 32-block regression with 128 source operations recorded 4,096 proof visits instead of 128 (projection-invariance-sharing-red-2.log). The first attempted fixture used an incorrect enum spelling and failed to compile; only the second run is behavioral red evidence.

Structural-artifact construction now derives one immutable proof for the exact source and shares it across block applications. Replacement and specialization still derive fresh facts. The regression checks both linear work and different facts for a changed source, while retaining the original owner’s facts. This changes compiler preparation, with no change to Modelica equations, AD rules, runtime value reuse, tolerances, or numerical acceptance. SPEC_0036 and SPEC_0043 §6a record the construction scope.

All 323 Solve IR, 100 Cranelift, and 503 solver tests pass, as does affected all-target/all-feature Clippy (projection-invariance-sharing-build-1.log). The four-model multibody-invariance-sharing-origin compares GyroscopicEffects, RollingWheel, and OvervoltageProtection high: three compared, one existing BevelGear1D reviewed exclusion, zero missing/deviating comparisons, and all 1,196 initialization channels high. The fixed multibody-invariance-sharing-canary retains every phase and band from multibody-reference-lineage-canary: nine compared high, no skips/missing/exclusions/nonidentifiability, and 175 high initialization channels. Receipts are invariance-sharing-origin-evidence-1.json and invariance-sharing-canary-delta.json.

Three alternating pairs using the exact ordinary MSL worker build measure:

Median lowering time drops from 8.4287 to 7.3694 seconds, a 12.57% reduction. All six complete traces are byte-identical. These are three unpinned pairs on a shared host, measuring lowering only. They establish no integration speedup or OMC performance win. invariance-sharing-ordinary-timing-series-1.json binds exact workers and traces. An earlier pilot omitted the harness’s test target from the build, changing dependency feature unification; its timing series is retained but superseded by these ordinary-worker measurements. The emitted Flat, DAE, structural DAE, and canonical Solve artifacts also match byte-for-byte (invariance-sharing-artifact-equality-1.json).

The initial diagnostic-output hypothesis was rejected: JSON-enabled profiling included roughly four seconds of requested artifact preparation/output after lowering, but the output-disabled control spent only 53 ms there. No worker shortcut or budget change was retained. The actual compiler fix addresses the repeated proof derivation. GyroscopicEffects passes the focused bounded attempt; its previous full-sweep timeout remains recorded until the next complete cohort establishes the result under full-run conditions. Combined verify quick/full remain pending.

IMS_Start has a reviewed reference-accuracy boundary

The convergence evidence below now supports a tracked comparator exclusion for IMS_Start. The model remains visible, receives zero strict-high or certification credit, and remains in the next complete cohort. No compiler equation, solver or comparator tolerance, official OMC trace, or baseline floor changes.

Quality-gate schema 5 records the policy migration against multibody-guarded-affine-full-11, with convergence evidence committed at 3d76411c1a41a1b27e6a0ecbf1cf3e204f0b47a4: strict-high remains 160, reviewed boundaries increase from 19 to 20, and the newly excluded model was non-high. These are classification counts on the retained failed sweep, not a new run or baseline promotion. The exclusion artifact’s SHA-256 is 30f7c38e58307d6af4f69b844512679ec860f367f88670481edad5318d7d29f8. The checked baseline retains all its previous measurements and certified model identities; only its gate version and new reference-boundary metadata change. Historical comparator, partial-classification, and promoted-baseline lineage evidence remain unchanged. Both baseline readers reject missing or forged boundary evidence; the remote-baseline resolver preserves all unrelated ratchets across version 4 to 5. The exclusion digest is checked against the exact tracked artifact.

Validation passes: 97 quality-gate tests, 169 harness library tests, 21 baseline-resolver tests, and affected all-target/all-feature Clippy. The final multibody-reference-lineage-canary retains every phase and band from multibody-trace-json-canary: nine compared high, zero skipped/missing/excluded or nonidentifiable traces, and all 175 initial-condition channels high. Evidence is ims-start-boundary-{migration-1,canary-delta}.json and ims-start-boundary-lineage-tests-1.log. Combined verify quick/full and the next complete cohort remain pending; this change does not establish release readiness.

The decoder regression remains the focused test for the actual transport bug. The recorded default/reference-refinement comparisons establish the numerical oracle boundary; they do not certify untested conditions or turn the model into a high-parity result. GyroscopicEffects’ compile-budget failure still requires a compiler fix, and the failed full sweep remains failed.

Full guarded-tearing sweep exposes trace transport and remaining failures

The complete multibody-guarded-affine-full-11 sweep at b0217592 fails its quality gate. It compares 161 models: 160 high, one near, 19 reviewed exclusions, and zero missing or nonidentifiable traces. IMC_DOL and IMC_Steinmetz newly compare high. IMS_Start newly simulates but has 690 high, 14 near, and six deviating channels out of 710. GyroscopicEffects loses its previous high result because Solve lowering exceeds its 10-second phase budget; its previous lowering time was 8.805 seconds. MultiBody therefore measures 21/42 high in this attempt, down from 22/42. The original failure remains recorded in guarded-affine-full-delta.json; no retry or partial snapshot replaces it.

IMS_Start revealed a transport defect before any equations needed changing. The solver’s JSON contains distinct left-limit and settled times 0.09999999999999999 and 0.1. The default JSON decoder rounds both to 0.1, although the voltage metadata correctly declares a continuous channel. The comparator then loses a side of the discontinuity and creates an interpolation ramp across the preceding output interval. The two reduced red tests prove both the changed timestamp bits and an artificial error integral of 3.50000000385 instead of approximately 3.85e-9 (trace-json-roundtrip-red-2.log). This is a decoder defect, not a reason to merge event coordinates or loosen comparison.

The workspace JSON dependency now enables float_roundtrip. The same shared transport preserves finite IR coefficients and literals as well as trace samples. SPEC_0033/0050 and SPEC_0036 record exact finite-value transport. Three regressions cover adjacent event times, the manufactured ramp, signed zero, subnormal values, and extreme finite values. All 187 DAE IR, 322 Solve IR, and 131 simulation facade tests pass, with affected Clippy clean (trace-json-roundtrip-green-1.log, trace-json-roundtrip-libraries-build-1.log).

Recomparing the unchanged IMS_Start files through the production loader and comparator removes four voltage-channel deviations, but does not close the counterexample: 690 channels remain high, 18 near, and two deviate. The remaining worst channels are aliases of a nearly zero summed current; OMC reaches about 9.24e-6 A while Rumoca stays near 1e-13 A. Their numerical and equation-level origin remains under investigation. No simulation tolerance, comparator threshold, or exclusion changed. The helper asserts the decoded adjacent timestamps remain distinct; its exact libraries and result are retained in trace-json-helper-libraries-1.json and ims-start-roundtrip-comparison-1.json.

The fixed multibody-trace-json-canary retains every phase and band from multibody-sparse-tearing-sensitivity-canary: nine compared high, zero skipped, missing, excluded, nonidentifiable, or deviating comparisons, and all 175 initial-condition channels high (trace-json-canary-delta.json). This completes focused validation of the shared decoder fix. IMS_Start’s remaining numerical differences, the GyroscopicEffects timeout, and combined verify quick/full remain open; no cohort baseline is promoted.

The same xtask-generated IMS_Start OMC executable was then evaluated at two higher numerical accuracies, changing only the copied initialization XML’s experiment tolerance. Executable, XML, sparsity inputs, CSV, and trace digests are recorded in ims-start-omc-refinement-{2,3}/receipt.json. The initial diagnostic bundle omitted OMC’s sparsity files and failed before initialization; the corrected bundles include those exact inputs. Both numerical runs finish successfully. Using the production comparator with the repaired decoder:

Rumoca’s worst bounded channel error against the two refined references is approximately 0.00256. The reference-to-reference comparison’s worst channel is der(aims.i_0_r) at 0.03731; all 735 channels meet the existing high threshold. MSL’s SpacePhasor component defines -m*zero.i = sum(i), and generated OMC equations 381/449 evolve the two zero-sequence currents from their inductor voltages. Together with the original summed-current traces, the convergence results identify an accuracy limitation of the default OMC reference. They provide no reason to make Rumoca reproduce its small spurious currents. This diagnostic does not replace the official reference or promote a cohort count. Reference-quality handling remains open, alongside the GyroscopicEffects compile-budget regression. Results are ims-start-refined-comparison-{2,3}.json and ims-start-omc-self-comparison-3.json.

Guard affine tearing and scale sensitivity checks by their direction

Affine projection now consumes a checked elimination layout derived from the exact block’s tearing and AD sparsity. RollingWheel’s 24-coordinate block reduces to five tear coordinates. The runtime retains sparse coefficients, the recovery relation, and the reduced factorization; changed conditioned coefficients revoke the cached factor. It solves the complete original right-hand side and recovers every coordinate, including corrections during mandatory original-equation refinement. Weak pivots or failed reduction retain the full implicit solve from the same arithmetic origin and tolerance. The original full AD Jacobian is still evaluated; this is not yet a compiler-composed reduced derivative kernel.

The AD pattern conservatively contains three future causal dependencies that structural tearing eliminated through invariant values. A replay at all 501 recorded RollingWheel coordinates found their coefficients exactly zero (inspect-affine-guards-result-1.json). This sample is not an all-state proof: the checked layout issues explicit guards, and the runtime checks fresh, unconditioned coefficients before every solve. Nonzero and nonfinite guards decline reduction, including a nonzero coefficient that conditioning would underflow to zero. No model names or frozen parameter assumptions select this path. SPEC_0036, SPEC_0043, and SPEC_0038 record these obligations.

The initial focused run exposed a BevelGear1D execution regression; it was retained as a failure despite that model’s existing comparison exclusion. Capturing its first failed sensitivity block showed a finite direction as large as 3.25e11, certified against scales from much smaller primal coordinates. Row 1191’s original JVP residual was -1.172149e-6: its ratio to the unchanged 1e-10 tolerance was 8.07 with the primal scale, but 0.0535 with the direction scale (bevel-sensitivity-scale-diagnosis-1.json). This identifies a sensitivity acceptance defect, not an invalid tearing layout.

A two-equation regression reproduces the defect independently of MSL: 17*a + 19*b = 23*s, 31*a + 7*b = 37*s, with analytic sensitivities da/ds = 271/235, db/ds = 42/235. Scaling the seed to 1e12 falsely failed the old check (sensitivity-scaling-red-2.log). The cached linearization now retains its original Jacobian and computes acceptance scales from each solved direction. The final original JVP and tolerance are unchanged. Positive and negative scaled directions pass; overflow and deliberately inconsistent JVPs still fail and restore incoming algebraic seeds. All temporary capture probes were removed before validation.

The final implementation passes 322 Solve IR and 503 solver tests, plus all-target/all-feature Clippy (sparse-tearing-sensitivity-build-1.log and sparse-tearing-sensitivity-origin-1.log). The three-model origin run restores BevelGear1D to sim_ok with its existing reviewed exclusion, while RollingWheel and OvervoltageProtection compare high: two compared, one excluded, zero missing or deviating comparisons, and all 229 initial-condition channels high. The fixed 20-model canary has no phase or band changes: nine compared high, zero skipped, missing, excluded, or nonidentifiable comparisons, and all 175 initial-condition channels high. Receipts are sparse-tearing-sensitivity-origin-delta.json and sparse-tearing-sensitivity-canary-delta.json, against the corresponding multibody-register-constants runs. These focused checks do not update the last complete cohort count or establish completion of combined verify quick/full.

Three alternating ordinary-worker timing pairs retain every candidate trace byte-for-byte equal to the compared origin trace, with identical canonical IR between candidate and control. Sim times in milliseconds were:

The medians are 73.600 and 74.726 ms, an observed 1.51% reduction with overlapping ranges. Three unpinned pairs on a shared host do not establish a reliable speed improvement. Preparation medians are 10.650 and 10.397 seconds, respectively; these are separate from Sim. The retained receipt is guarded-tearing-timing-series-1.json. The analysis helper initially expected IR files in the focused harness directory, which does not emit them; its bookkeeping repair reused the completed first sample without rerunning it.

Exact-worker perf samples confirm the guarded solve is active in all three candidate runs. The control profiles retain faer’s full sparse numeric factorization; the candidate profiles sample TornNewtonCache::solve_scaled and the existing full rumoca_projection_871_application. The sparse reduced factorization therefore fixes the blanket dispatch bypass, while removing the full derivative assembly remains a separate compiler task. Logs and symbol maps are retained in guarded-tearing-final-* and guarded-tearing-perf-leaves-1.log. RollingWheel is still slower than the pinned OMC measurements; this change makes no OMC speed or end-to-end performance claim.

Why affine projection previously bypassed tearing

Commit 82630a4d placed affine projection before tearing to repair stale switching coordinates: a residual below tolerance could still leave a voltage on the wrong side of zero. The affine path computes A*x = -F(0) and refines against original residuals. Later commits c4789fce and 3a7c0b12 preserve small coordinates beside large offsets. These guarantees do not inherently require the full block, but the existing reduced Newton path uses finite differences and different acceptance rules; reordering dispatch alone does not establish equivalent behavior.

The current RollingWheel block has 24 unknowns, five tears, and 19 exact back-substitution steps. A private checked-IR replay restored all 916 solver coordinates through the trace’s exact LoadY output projections at all 501 recorded times. Every substitution sweep completed, both from recorded tears and from zero tears. At recorded tears, the maximum absolute recovered-value change was 4.55e-13 and the maximum reduced residual was 2.16e-13. This rejects an unusable isolator at those sampled points as the reason for the bypass; it does not prove usable pivots at every integrator stage or validate a reduced solver. The evidence is torn-sweep-trajectory-result-1.json, with its helper and input mapping retained under .git/multibody-campaign/rolling-wheel/.

OMC’s generated linear system 731 has six unknowns and an analytical Jacobian. The implementation target identified by this inspection was an AD-derived reduced affine system from the issued substitution schedule, preserving original-equation refinement and recovered-coordinate accuracy. An unusable reduced pivot must retain the original implicit solve. The inspection itself made no dispatch change or new timing claim; the latest complete MSL result remains the run below.

Invalidate native constant facts on register writes

A checked Solve program exposed a native indexing counterexample: it sets register 0 to index 1, overwrites that register from parameter input 2, then indexes {10,20}. Cranelift’s constant-index shortcut retained the old fact and selected 10; the reference evaluator correctly returned 20. The actual red test records that exact native/reference mismatch in register-constants-red-1.log. The first divergence is native constant tracking, downstream of a valid register-flow certificate.

Constant facts now belong to register versions. The emitter computes a new fact from an operation’s input versions, emits the complete operation, removes old facts throughout its destination range, and installs the new fact if proven. Tensor invalidation visits existing facts without expanding tensor extents. The same rule covers dynamic writes and in-place arithmetic. SPEC_0036 and SPEC_0043 §6a record the lifetime; MLS §10.5 supplies the indexed array semantics. Tests cover scalar overwrite, both destinations of a tensor write, and in-place addition. All 100 Cranelift tests and affected Clippy pass.

RollingWheel’s complete IR and trace are byte-identical to the preceding projection-reuse build. Its one diagnostic run took 73.695 ms; that single sample is not a new performance claim (register-constants-profile-evidence-1.json). The origin check retains two high comparisons, the existing BevelGear1D reviewed exclusion, no missing traces, and 229 high initialization channels. The fixed canary retains nine comparisons, all high, no skipped or missing traces, and 175 high initialization channels. Every phase and band is unchanged (register-constants-origin-delta.json, register-constants-canary-delta.json).

The complete sweep at 0c22de12 passes in target/msl/multibody-register-constants-full-11: 159/566 strict-high (28.09%), 159 compared, 19 reviewed exclusions, no missing or non-identifiable traces, no deviation channels, and all 20,964 initialization channels high. All phase and band results match the previous projection-reuse sweep, leaving MultiBody at 22/42 high (register-constants-full-delta.json).

A subsequent allocation experiment reused the solved Newton vector while unscaling it. All 75 projection tests and Clippy passed, but three alternating ordinary worker pairs measured 73.900 ms versus 74.261 ms median with overlapping samples and identical complete traces. The 0.49% difference did not establish a clear speed improvement, so the experiment was removed and the ordinary worker rebuilt (owned-delta-timing-summary-1.json).

The exact worker dependency chain enables faer’s Rayon feature, but replay of the issued 24-by-24, 96-entry pattern selects simplicial LU. That implementation’s numeric factorization and triangular solves run serially; forcing sequential configuration cannot remove threaded work from this block (faer-issued-parallelism-evidence-1.json). The larger unresolved issue remains the full coupled system and its repeated evaluations, compared with OMC’s six-unknown generated linear system.

Reuse primal work across projection-Jacobian colors

The complete projection application now issues seed-invariance facts using the existing register-source checker and whole-program repeatability proof. Every write replaces its destination version’s evidence. Cranelift retains certified results as native SSA values after their first ordered execution and reuses them across subsequent colors in that call. Tensor operations stay complete; mixed seed loads remain live. Source replacement rederives the facts, each call uses fresh coordinates, and native failure prevents matrix publication. SPEC_0036 and SPEC_0043 §6a define this boundary.

The work-count regression first failed because a seed-independent sine ran twice for two colors. It now runs once per call and recomputes for changed Y/P/time. Controls cover overwritten registers, mixed tensor tangent lanes, owner identity, output placements, table validation, and a singular typed tensor solve followed by successful calls at fresh parameters. All 318 Solve IR tests and 96 initial Cranelift tests pass, as does the added native-failure test (415 distinct tests total). Affected all-target/all-feature Clippy passes.

For RollingWheel, the issued proof certifies 378 of the hottest program’s 536 operations, including 26 matrix products, 12 cross products, and four pure calls. This is stronger evidence than the earlier output-dependency census; unproven operations still execute in every selected color. Three alternating ordinary msl-fast worker pairs measure 75.426 ms median simulation time versus 81.620 ms for the preserved control, a 7.59% reduction. All Flat, DAE, structural DAE, Solve, and complete trace artifacts are byte-identical across the six runs. Preparation increases from 9.899 s to 10.945 s median. These unpinned samples show a simulation improvement, not an end-to-end speedup or a win over OMC. Receipts are projection-reuse-alternating-evidence-1.json, projection-reuse-timing-summary-1.json, and projection-reuse-issued-counts-1.json; the exact workers and native perf maps are retained with the profiles.

multibody-projection-reuse-origin retains two high comparisons, one existing BevelGear1D reviewed exclusion, no missing traces, and 229 high initialization channels. The fixed canary multibody-projection-reuse-canary retains nine comparisons, all high, no skips or missing traces, and 175 high initialization channels. Every phase and band is unchanged from the corresponding derivative tensor runs (projection-reuse-origin-delta.json and projection-reuse-canary-delta.json).

The full sweep at 5d1a8f06fd7658892f5aa4e4573c103b73885337 passes in target/msl/multibody-projection-reuse-full-11: 159/566 strict-high (28.09%), all 159 compared models high, 19 reviewed exclusions, no missing or non-identifiable traces, no deviations, and all 20,964 initialization channels high. MultiBody remains 22/42 high. Every phase and band is unchanged from multibody-derivative-tensor-full-11 (projection-reuse-full-delta.json). Combined verify quick/verify full and the older coverage ratchet repair remain outstanding; this full MSL result alone does not establish PR readiness.

Share derivative tensor definitions across component projections

The reduced case der(x)=A*x; z=sum(der(x)) exposed duplicate tensor computation in Solve lowering. Both derivative components used the same matched definition, but their scalar-view domain points created different expression contexts. The regression initially emitted two complete matrix products. The definition’s construction-issued binder metadata now distinguishes a real binding dependency from an output projection, reducing that case to one product. A binder-dependent control retains separate products and correct values for both tested parameter/state tuples. SPEC_0032 §2 records the boundary.

All 141 Solve tests and 605 core tests pass, together with affected Clippy and formatting checks. OMC agrees exactly with all four initial-equation values used by the reduced controls (derivative-tensor-sharing-omc-1/receipt.json); this is an equation check, not a trajectory comparison. The focused originating run multibody-derivative-tensor-origin retains two high comparisons, one existing reviewed exclusion, no missing traces, and 229 high initialization channels. The fixed 20-model canary multibody-derivative-tensor-canary retains all previous phase and band results: nine compared models, all high, no skips or missing traces, and 175 high initialization channels. Receipts are derivative-tensor-origin-delta.json and derivative-tensor-canary-delta.json.

The complete 566-model sweep at e880636bf2feef099de096d4d998b6190cf30632 passes in target/msl/multibody-derivative-tensor-full-11: 159 strict-high models (28.09%), 159 compared, 19 reviewed exclusions, no missing traces or deviations, and 20,964 high initialization channels. MultiBody remains 22/42 strict-high. Every phase and comparison band matches the previous passing multibody-invariant-binding-full-11 sweep (derivative-tensor-full-delta.json).

This fix does not change RollingWheel’s generated kernel. Its Flat, DAE, structural DAE, Solve, and complete trace remain byte-identical to the control (derivative-tensor-sharing-profile-evidence-1.json). The fresh single run is 86.469 ms versus 79.726 ms for the control; unchanged executable IR and single samples do not establish a speed effect. Checked replay identifies the hot Jacobian program as Body’s a_0=der(v_0): its primal has 392 operations and its directional program 536, executed in six of the block’s nine colors. The remaining expansion must be traced through structural differentiation.

A diagnostic using the canonical seed/effect dependency authority finds 435 of that program’s 535 result-producing operations independent of seeds and free of recorded opaque effects, including 31 of 38 matrix products and 16 of 20 cross products. The program executes for six colors at one numerical coordinate (derivative-tensor-projection-seed-independence-1.json). Reusing first-color calculations is therefore a concrete next investigation. These output-dependency counts are not an execution-reuse certificate: any retained implementation must also prove operation/version ownership, failure behavior, scratch lifetime, and unchanged coordinates before omitting evaluations.

Three temporary cache experiments produced no retained optimization: bypassing all pure-call result caches, bypassing small arithmetic bodies, and inlining small cache copies. Their receipts explicitly retain the single-sample and build-configuration limitations; they do not establish an optimal cache policy. All probes were removed and the ordinary worker rebuilt before validation.

Structural functions retained in symbolic bindings

The full sweep at 617e315a2f3986dceddb1d3096b4ddc4738e9d29 exposed a regression missed by the fixed canary: six previously high electrical models and five previously excluded simulation completions fail Flatten. The run reports 153/566 strict-high, all 153 compared models high, 14 reviewed exclusions, and zero missing, nonidentifiable, or deviating comparisons; it fails the quality gate (zero-parent-full-delta.json). Thirty phase outcomes change, mostly from the same missing structural function. GearConstraint also changes from structural refusal to a Solve timeout; it remains unsupported.

The electrical failures share MultiDelta/MultiStar’s declaration mSystems=numberOfSymmetricBaseSystems(m). Function precollection scans only the evaluated instance binding, now a literal, while Flatten correctly retains the symbolic source call. The integer evaluator therefore lacks the function needed to resolve dimensions and connection-loop bounds. This is a missing dependency in Flatten preparation, not permission to freeze mutable Real parameter bindings again (SPEC_0007/0040; MLS §4.4.4, §8.3.3, §12.4).

Precollection now consumes the same retained binding source as variable flattening, preserving the existing transitive callee discovery. A reduced imported-function regression with two differently modified instances fails before the change and passes afterward, retaining symbolic calls while producing three and five array coordinates and eight equations (structural-binding-functions-{red,focused}-1.log). Pinned OMC produces the same eight exact observable values in all six output rows (structural-binding-functions-omc-1/receipt.json). All 631 flattening tests, 602 compiler-core tests, and affected Clippy pass (structural-binding-functions-broad-build-1.log).

The eleven-model originating run still fails after that preparation repair: all eleven pass Flatten, but five now reject nonliteral array extents during DAE construction and six reject recursive functions in the executable pure-call graph. It compares zero traces and has parity unmeasured (structural-binding-functions-origin-1.log). Translation-time evaluations must remain available to structural consumers without freezing changeable parent bindings or requiring compile-time recursion to execute at runtime. Neither this originating run nor the earlier failed full sweep is passing evidence.

A second reduced case reproduces the next boundary exactly: a recursive pure Integer function computes two instances’ fixed array widths, three and five; fill(1.0, width) then fails DAE construction with a nonliteral extent (invariant-binding-recursion-red-1.log). Under SPEC_0007’s FLAT-C01 contract, Flat finalization now specializes scalar bindings only after following exact instance/declaration references through every fixed, non-changeable parent. The shared constant interpreter owns function evaluation, including recursion, purity, assertions, and evaluation limits. It receives no unproven global parameter values. Failed or unsettled evaluations retain their bindings; array-valued bindings remain symbolic. Existing function pruning removes the now-unreachable compile-time helper.

The recursive fixture now compiles and simulates both arrays and their integrals. Four focused tests pass, including a transitive final-alias chain whose legal parent override changes its output, and refusal to fold impure or fixed=false dependencies (invariant-binding-controls-1.log). Pinned OMC produces the same literal widths and all ten analytic observable values in six rows, with maximum absolute error 1.12e-16 (invariant-binding-recursion-omc-1/receipt.json). All 631 flattening tests, 605 compiler-core tests, and affected Clippy pass (invariant-binding-broad-build-1.log).

The eleven-model originating run now passes: all eleven simulate; six compare high, including all 2,205 initialization channels, and five retain their previous reviewed exclusions. There are no missing or deviating comparisons. All eleven bands match their entries in the last passing full sweep at 59cee76f (invariant-binding-origin-delta.json). The prior failed originating run had no measured bands; the receipt distinguishes its phase outcomes from that complete baseline’s band evidence. The fixed 20-model canary also passes: all phases and bands remain unchanged, nine models compare high with 175 high initialization channels, and none are skipped, missing, or deviating (invariant-binding-canary-delta.json).

The complete multibody-invariant-binding-full-11 sweep at cb9920bcd84b57287f545a0110364d2af242ba0f passes: 159/566 strict-high (28.09%), all 159 compared models high, 19 reviewed exclusions, zero missing, nonidentifiable, or deviating comparisons, and all 20,964 initialization channels high. MultiBody remains 22/42 high. Every simulation band matches the last passing full sweep at 59cee76f; the only phase-field difference is an error-code change from ED020 to ED019 for the already unsupported PolyphaseRectifier (invariant-binding-full-delta.json). The comparison to the failed 617e315a sweep separately records restoration of its regressions (invariant-binding-full-regression-repair-delta.json). This restores cohort coverage; it establishes no RollingWheel speed improvement.

RollingWheel kernel measurements after coverage recovery

The ordinary worker at cb9920bc runs RollingWheel in 79.894 ms under the development benchmark profile. Its Flat, DAE, structural DAE, Solve, and full trace are byte-identical to the previous control (invariant-binding-profile-evidence-1.json). The 74-sample perf capture still includes sparse factorization, triangular solves, and Jacobian execution; it is a hotspot diagnostic, not a precise cost breakdown.

A controlled kernel comparison does not support replacing the current sparse solve with dense LU. The initial experiment used the 74-entry numerical union of three captured matrices. Checked replay of the saved Solve model reconstructs the actual 96-entry projection pattern; repeating with that exact pattern gives median costs of 3.195 microseconds for the current sparse path and 3.420 for dense LU, about 7.1% slower. Each median covers seven alternating rounds of 3,000 changing-matrix solves after a warmup round (dense-crossover-issued-measurement-1.json, dense-crossover-pattern-audit-1.json). This measures those kernels, not a whole-simulation speed ratio. The selection policy is unchanged and the temporary measurement code has been removed.

The previous elimination experiment did handle its zero pivot by retaining it in an eight-variable simultaneous system; it did not merely fall back to the full solve. Mapping the same captured matrices to the current five-tear plan finds no zero causal pivots or nonzero future causal dependencies at any of the three captured coordinates (affine-elimination-pivot-recheck-1.json). This does not prove nonzero pivots throughout a trajectory or establish a faster implementation. The production affine executor still solves all 24 coordinates.

Invariant-zero tensor incidence and declaration binding repair

The RollingWheel investigation reached a second producer defect: scalar dependency projection reports off-diagonal inertia dependencies even when the selected coefficient is a fixed, non-tunable literal zero. Under SPEC_0032 §2 and MLS §10.6.4, the projection now proves such zeros through the original literal, parameter, array, concatenation, and signed-expression owners. It omits only the multiplied direct Real coordinate’s incidence; the canonical tensor expression and numerical evaluation remain intact. Tunable parents, initialization unknowns, and arbitrary expressions with domain obligations do not qualify. The reduced matrix dependency test fails before the change while these negative controls already pass (zero-coefficient-red-1.log). All 46 evaluation, 139 Solve, and 170 structural tests pass, with affected Clippy.

The first ordinary RollingWheel diagnostic changes the normalized plan from seven tears to five and replaces the zero causal pivot with diagonal inertia coefficients. Flat, DAE, structural DAE, and the complete trace are byte-identical to the final-modifier control. Its 78.924 ms runtime is not a speedup: the affine executor still solves the full 24-coordinate system. The receipt is zero-coefficient-profile-evidence-1.json.

A runtime parameter-override control then exposed an existing correctness counterexample. In ParentCoefficientProbe, final parameter Real offDiagonal=parent was frozen to zero during instantiation because parent defaults to zero. With I=[2,offDiagonal;offDiagonal,2] and I*x={1,2}, the legal override parent=0.5 must produce x={1/3.75,3.5/3.75}. Rumoca instead kept {0.5,1}. The archived ordinary worker also emits the wrong literal Flat binding, ruling out the new incidence proof or the numerical solver as the first divergent owner (zero-coefficient-baseline-flat-1/).

Instantiation now retains the original declaration binding when evaluating a different value for structural queries. Flattening consumes that retained symbolic source using the existing declaration scope. This preserves the binding equation and changeable-parent dependency under MLS §4.4.4 and §7.2.6. The direct instantiation regression fails before this repair (zero-parent-phase-red-2.log); all 224 instantiation, 631 flattening, and 601 compiler-core tests now pass, including the actual runtime override and Flat binding regressions. Affected-package Clippy passes (zero-parent-focused-1.log, zero-parent-broad-build-1.log). Pinned OMC agrees with all four reduced default/override cases to within 5.6e-17 (zero-coefficient-omc-1/receipt.json).

The combined ordinary RollingWheel diagnostic retains the same complete trace and five-tear plan at 78.591 ms. Its sole Flat change preserves the declaration binding world.groundLength_v=world.groundLength_u instead of literal 4 (zero-parent-flat-delta-1.json, zero-parent-profile-evidence-1.json). Tier 1 originating comparison retains RollingWheel and OvervoltageProtection high: two compared models, 229 high initialization channels, one reviewed BevelGear exclusion, and no missing or deviating comparisons. The fixed twenty-model canary retains every phase and band: nine compared models, all nine high, 175 high initialization channels, and zero skipped, missing, nonidentifiable, or deviating comparisons. Receipts are zero-parent-{origin,canary}-delta.json.

Affine elimination experiment and final-modifier investigation

The complete multibody-tearing-choice-full-11 sweep at 2892128a7e9337989b455fe2911e7c6f52bffe51 passes with every phase and band unchanged: 159/566 strict-high, 159 compared models, 19 reviewed exclusions, zero missing/nonidentifiable/deviating comparisons, and 20,964 high initialization channels. MultiBody remains 22/42 high. The receipt is rolling-wheel/tearing-choice-full-delta.json.

Three captured RollingWheel matrices at t=0, 0.5005823461, and 2.0053747285 expose one zero causal pivot in the seven-tear plan: residual 782, solver coordinate 43 (body.z_a[2]). Its canonical tensor program is the first component of Body.mo line 261, frame_a.t = I*z_a + cross(w_a, I*w_a) + cross(r_CM, frame_a.f). The coefficient is I_21, set to zero with a final modifier in RollingWheel.mo. The other causal diagonals are plus or minus one, with no upper causal coefficients. The capture preserves every IR and complete trace byte; all probes were removed (affine-matrices-1.json, affine-matrix-capture-receipt-1.json).

A checked Schur-elimination experiment retained unsuitable pivots in the simultaneous system, recovering all original coordinates under unchanged conditioning, residual checks, and mandatory refinement. Two 24-coordinate cycle regressions failed before it; the all-unsuitable-pivots control already passed. All three subsequently passed, along with four construction controls and all 1,006 affected library tests. Clippy passed after removing two redundant test-only copies. Logs are affine-elimination-red-1.log, affine-elimination-libraries-1.log, and affine-elimination-clippy-build-1.log.

Its ordinary run took 82.948 ms against an immediate archived control at 77.444 ms, 7.11% slower (affine-elimination-profile-delta-1.json). Perf identified factor construction as a major cost. Skipping zero-coefficient products and retaining the existing small-matrix policy passed all 78 projection tests and Clippy, but still took 80.323 ms. The originating comparison retained both RollingWheel and OvervoltageProtection high: two compared models, 229 high initialization channels, one reviewed BevelGear exclusion, and no missing or deviating comparisons (affine-elimination-origin-delta.json). Compiler IR bytes were unchanged; trace bits changed. Neither candidate demonstrated a speedup. Both implementations, their tests, and proposed contract extensions were removed and preserved in affine-elimination-first-candidate-1.patch, affine-elimination-rejected-2.patch, and the associated new-file archives. No canary or cohort claim is made for this rejected runtime experiment.

The earlier divergent compiler layer is now under investigation. OMC retains body.I_21 as a non-changeable calculated parameter, but Rumoca’s Flat flags contain only the declaration-final body.I, omitting modification-final I_21, I_31, I_32, and r_CM. Instantiation records declaration finality while ignoring the accepted modifier’s final prefix. Three reduced tests fail for a literal modifier, an each-final component array, and a symbolic parent binding; the declaration-final and ordinary-sibling controls behave as expected (final-modifier-red-1.log). Final modifiers must retain their source bindings, including dependencies on changeable parent parameters, under MLS §7.2.6. The candidate now carries effective finality through instance construction and nested scopes. All 223 instantiation and 598 compiler-core tests pass, as does affected-package Clippy. Pinned OMC agrees on the exact reduced model and on a tensor regression with source modifications and a legal runtime parent override to gain=4; all three trajectories match their analytic solutions within 2.3e-16 in OMC. The artifacts are final-modifier-omc-1/ and final-tensor-omc-1/receipt.json; logs are final-modifier-focused-1.log, final-modifier-core-2.log, and final-modifier-core-build-1.log.

The ordinary RollingWheel diagnostic now records the effective final flags in Flat and non-tunable attributes in DAE, but Solve and the complete trace remain byte-identical. Its 78.849 ms runtime is not a speedup (final-modifier-profile-evidence-1.json). Tier 1 originating comparison retains two high models and 229 high initialization channels, with one reviewed BevelGear exclusion and zero missing/nonidentifiable/deviating comparisons. The fixed twenty-model canary retains every phase and band: nine high compared models, 175 high initialization channels, and no skipped, missing, nonidentifiable, or deviating comparisons. Receipts are final-modifier-{origin,canary}-delta.json. The full multibody-final-modifier-full-11 sweep at 59cee76f19ec6fc0f97121e45b506d7fa6086d8b then retained all phases and bands: 159/566 strict-high, all 159 compared models high, 19 reviewed exclusions, zero missing/nonidentifiable/deviating comparisons, and 20,964 high initialization channels. MultiBody remains 22/42 high. Its receipt is final-modifier-full-delta.json; tracked source was clean at launch.

Rejected native colored-application experiment

A native batch consumed each issued projection color and complete source program in order, sharing immutable Y/P loads and scattering the checked output placements in one native call. Eleven focused numerical/ownership/failure tests and Clippy passed. The ordinary RollingWheel run took 78.378 ms versus an immediate archived control at 78.045 ms. Every compiler artifact and complete trace remained byte-identical (projection-batch-profile-delta-1.json). The hot batch occupied 126,064 native bytes and remained a leading perf leaf; removing Rust dispatch did not demonstrate a speedup. The implementation was removed and archived in projection-batch-rejected-1.patch and projection-batch-rejected-source-1.rs. No MSL gate or parity credit is claimed for this experiment.

Its first standalone package test exposed an existing fixture wiring defect: Cranelift tests use StructuralPattern::from_row_dependencies, but only another package’s development dependency enabled pattern-fixtures. The adapter now explicitly requests that feature in its own development dependencies. Production dependencies and the checked pattern authority remain unchanged. The original failure is projection-batch-focused-1.log. All 94 standalone Cranelift library tests and Clippy pass after the fixture fix, and the ordinary worker has been rebuilt without the rejected batch (cranelift-fixture-final-1.log).

Current compiler work: compare complete tearing candidates

RollingWheel’s 24-coordinate force/torque/acceleration block exposes a structural cost defect: prioritizing immediate causal unlocks chooses twelve tears, then exact-assignment normalization promotes one more. A degree-first candidate needs seven. The source incidence graph is retained in the regression test and rolling-wheel/tearing-graph-heuristic-diagnostic-1.json. The probes preserve all compiler artifacts and trace bytes. Normalization is not the main cause; transitive-unlock lookahead also produces twelve tears. OMC’s generated linear system 731 has six unknowns, though its coordinates differ from Rumoca’s.

Under SPEC_0007/0040 STRUCT-T06 and MLS §8.3.1, the structural producer now compares two complete deterministic candidates and selects fewer tears, retaining the original on ties. Both preserve the exact causal restrictions and complete equation/unknown partitions. No source equation, tensor owner, runtime tolerance, or model-specific rule changes. The graph regression fails before the change (tearing-choice-red-1.log); partition, restricted-candidate, and input-order controls pass with the structural suite and Clippy (tearing-choice-focused-1.log, tearing-choice-focused-build-2.log).

The actual normalized Solve block changes from thirteen tears to seven while Flat, DAE, structural DAE, and the complete trace remain byte-identical (tearing-choice-profile-evidence-1.json). The ordinary msl-fast diagnostic takes 78.080 ms. This is a compiler plan improvement, not a demonstrated runtime speedup: the affine executor still solves the full 24-coordinate matrix.

Tier 1 multibody-tearing-choice-origin retains RollingWheel and OvervoltageProtection high: two compared models, 229 high initialization channels, one reviewed BevelGear exclusion, and zero missing or deviating comparisons. multibody-tearing-choice-canary retains every phase and band of the fixed twenty-model list: nine compared models, 175 high initialization channels, and zero skipped, missing, nonidentifiable, or deviating comparisons. The receipts are tearing-choice-{origin,canary}-delta.json.

A third candidate prioritizing branded derivative coordinates passed 171 tests, Clippy, the originating list, and the canary, but did not reduce this actual normalized block further: derivative aliases are algebraic coordinates at this stage. Its 77.516 ms diagnostic still has seven tears and an identical complete trace. The speculative extra candidate was removed; its patch, tests, and measurements remain in tearing-derivative-preference-experiment-1.patch, tearing-derivative-choice-focused-build-2.log, and tearing-derivative-choice-profile-evidence-1.json. No five-coordinate runtime claim follows from its isolated graph result. The final two-candidate source is restored exactly; its completed cohort evidence is recorded above.

Current cohort and RollingWheel measurement

The complete multibody-bdf-step-budget-full-11 sweep at commit 6e3b5c2d3a415db5d1191c7208e9c2968697d2ad passes: 159/566 strict-high models (28.09%), including 22/42 MultiBody. All 159 compared models and 20,964 initialization channels are high. There are 19 reviewed exclusions and zero missing, typed nonidentifiable, near, or deviating comparisons. Every previously high model is retained; IMC_YD additionally becomes high after the step recovery budget correction. BevelGear1D remains a reviewed exclusion, never a high result. The tracked source was clean; the foreign untracked communication file accounts for the harness dirty flag. The receipt is bdf-step-budget-full-delta.json.

RollingWheel remains slower than OMC. At that same commit, the ordinary release worker takes 71.411 ms versus an immediate msl-fast control at 77.773 ms. This 8.18% difference is a paired measurement, not repeated medians; release already uses ThinLTO and no build profile was changed. All Flat, DAE, structural DAE, Solve, and complete trace bytes are identical (release-lto-profile-delta-1.json).

OMC’s reported 16.582 ms simulation median excludes its solver timer. Inspection of solver_main.c at pinned OMC revision a96aa1a682c463b0fd2d285b486c09a8b7fe496d establishes the subtraction used by finishSimulation. Per-run sums over the archived 40 runs give 17.722 ms for simulation plus solver, 18.924 ms for total minus output, and 18.236 ms after also subtracting initialization and preinitialization. Total including CSV output is 50.655 ms. Those scopes differ from Rumoca’s prepared run with in-memory output; none establishes a Rumoca speed advantage. The receipt is omc-timer-accounting-1.json.

A removed 40-run timer probe has an 80.721 ms median, including instrumentation overhead. Numerical advance takes 52.475 ms median. Inclusive accumulated timings identify the 24-coordinate affine block at 38.306 ms per run, including 14.474 ms for Jacobian evaluation and 10.422 ms for linear solves. Timings include small build/initialization contributions and nested categories must not be summed. All repeated traces and compiler artifacts match the ordinary release byte-for-byte. Exact executable/JIT maps, perf leaf samples, source restoration hashes, and the measurement limits are retained in release-cost-repeat-profile-1 and release-cost-repeat-summary-1.json.

An experiment specializing AD further to each color passed 1,043 focused tests and affected-package Clippy. Its complete compiler artifacts and trace were byte-identical, but an immediate ordinary-worker pair took 82.120 ms versus 77.854 ms for the archived baseline (5.48% slower). Preparation took 9.677 s versus 9.743 s. No performance improvement was demonstrated, so the experiment and its proposed contract extension were removed. The exact patch, both worker executables, tests, and receipts remain in color-seed-domain-experiment-1.patch, color-seed-domain-focused-{1,2,3}.log, and color-seed-domain-profile-delta-1.json; the second focused attempt only exposed a test constructor unavailable under that feature set. This is rejected performance evidence, with no additional MSL coverage claim.

Latest focused work: distinguish step recovery from cumulative statistics

The full multibody-bdf-convergence-history-full-11 comparison at clean commit 6f203e1cc5e809229ac25a554c742beb1766e7e7 failed its quality gate. Of 566 targets, 158 were compared: 157 strict-high, one near (BevelGear1D, including one deviating channel), 18 reviewed exclusions, and zero missing or typed nonidentifiable traces. All 20,900 initialization channels were high. OvervoltageProtection regressed from strict-high to nonlinear-solver exhaustion; BevelGear1D previously refused a projection sensitivity and now finishes with an unacceptable trajectory. Neither is a performance success. GyroscopicEffects completes high in this full run; its earlier focused timeout remains recorded. Inverse_sh_TX changes from Flatten EF015 to ToDae ED019 without a compiler source change. The complete delta is retained in rolling-wheel/bdf-convergence-history-full-delta.json.

The electrical failure is at the numerical dependency, not a changed equation: the BDF failure counter accumulates 51 failures across 50 distinct accepted-step indices, with at most two failures in any one step. It aborts after 804 accepted steps at t=0.3748128288461084. The step limit is incorrectly checked against lifetime statistics. This remains present in upstream Diffsol commit 7036380f908dbd93baa4253d2e0a34aa115cbbb5; the frozen source and diagnostic receipts are diffsol-upstream-bdf.rs and bdf-failure-count-probe-summary.json.

The direct constant-derivative regression with injected recoverable failures fails before the correction, while its persistent-failure negative control already passes (bdf-failure-budget-red-2.log; red-1 only exposed a missing test type annotation). BDF now checks the unchanged configured recovery limit against failures in its current step, preserving lifetime statistics and rejection of an unresolved step. All ten dependency tests and all-target, all-feature Clippy pass (bdf-failure-budget-focused-1.log). The ordinary multibody-bdf-step-budget-origin run restores OvervoltageProtection: all 45 trajectory and initialization channels are high. RollingWheel retains all 184 high channels and a byte-identical trace. These are two comparisons, one reviewed exclusion, and zero missing or deviating comparisons. The fixed multibody-bdf-step-budget-canary preserves every preceding phase and band: nine comparisons, 175 high initialization channels, and zero skipped, missing, nonidentifiable, or deviating comparisons (bdf-step-budget-canary-delta.json). The tracked-exclusion validation test passes; the subsequent complete cohort result is recorded above.

BevelGear’s reference was examined independently of Rumoca: the unchanged, xtask-generated OMC executable produces revolute1.phi(1) of 72.3016880708 rad at its default 1e-6 tolerance, 48.8580301756 at 1e-9, and 48.1459606592 at 1e-12. Each diagnostic uses a separate copy of its initialization XML with only DefaultExperiment.tolerance changed, supplied with -f; the tracked reference and Rumoca tolerances remain unchanged. At 1e-12, perturbing only the fixed initial revolute1.phi by 1e-8 rad changes the endpoint to 41.0327222753 rad. The first diagnostic incorrectly used -override=tolerance=...; OMC warned that this variable did not exist, so those runs provide no tolerance-refinement evidence. Both attempts are retained as rolling-wheel/bevel-omc-convergence-{1,2}/receipt.json with the exact commands and executable, XML, and trace digests. The OMC executable SHA-256 is 444907bed49f3347e26913253c351bbd73c4e086e72a3c481fbcfdb4247325cf; the original XML is 03427ae6499f2835cd528f80fa6ad0ddf8a92708f90aa2118fb74fc6daf7f139. The 1e-12 unperturbed trace used by the derivative comparison is a273bce69959d7dada8f34d653a91f26cd1a9c4cefc03e00c943c06247ff78ac.

To reject a different-equations explanation, an isolated, removed adapter probe calls the ordinary component derivative handle at all 502 output points from the 1e-12 OMC trace. State order comes from the actual Solve layout; the extra gear-output angle derivative is the connected inertia2.w. All 5,020 derivatives agree: maximum absolute error is 6.8453e-8, and maximum abs(error)/max(1,abs(reference)) is 1.4462e-11. Every Flat, DAE, structural DAE, and Solve artifact is byte-identical to the preceding BevelGear worker. bevel-point-probe-summary-1.json retains the worst values and their times; bevel-omc-points-1.json binds the exact reference CSV. The adapter is restored to its archived source hash before ordinary validation.

The reviewed exclusion records this demonstrated default-reference limitation. It does not claim a proved positive infinite-time Lyapunov exponent, invariant refinement, or high trajectory parity. The strong initial-condition sensitivity and OMC self-divergence make a single default OMC trajectory non-identifying; the model remains outside the strict-high count, with invariant and sensitivity obligations outstanding. No compiler, runtime, tolerance, or model-name branch is added to obtain agreement.

Prior focused work: consume BDF correction history at the corrected point

Diffsol’s backtracking Newton path never populated the correction-norm history used by its convergence-rate estimate. Three direct dependency regressions fail against that path (bdf-convergence-history-red-2.log): it repeats residual evaluation for affine solves, oversolves a rapidly contracting nonlinear root, and fails to reject a rate that cannot meet the iteration budget. The earlier red-1 attempt only exposed test-import errors.

A history-only change is incorrect: a successful line-search trial has computed the next Newton correction but has not applied it. The negative control in bdf-convergence-naive-history-red.log returns 0.0098039 for an exact root of zero when the requested nonlinear bound is 0.0002. The corrected iteration consumes that pending correction and its norm together, applying the correction before success. Armijo backtracking, tolerances, iteration budgets, and the existing Jacobian/timestep estimate-reset policy remain unchanged. Frozen upstream commit 7036380f908dbd93baa4253d2e0a34aa115cbbb5 retains the original behavior; the dependency patch and provenance are recorded in vendor/diffsol/RUMOCA_PATCH.md.

All eight direct dependency tests, 162 simulation tests, both sampled-integral event-entry regressions, and affected-package all-target/all-feature Clippy pass (bdf-convergence-history-focused-{1,2}.log). The ordinary worker 1e90fe17cb1338e608a6de49b4a12fa635107ecf0818b4132b7017baa9ca6a6a takes 79.153 ms against the immediate archived 4b4dc9c5… recheck at 98.345 ms: 19.51% improvement for that pair. Compiler artifacts are identical; trajectory bits change with the nonlinear iteration (bdf-convergence-history-profile-delta-1.json). The separate 40-run profile has a 79.444 ms median (78.917–83.315 ms), with every repeated trace bit-identical. RHS calls fall from 2,248 to 1,560; accepted steps change from 1,054 to 1,051. The new run has 58 error-test failures, two nonlinear failures, 109 linear setups, and 17 complete matrix evaluations. These remain numerical statistics, not parity evidence. Diagnostic probes are removed and their two source files restored exactly before ordinary validation.

The canary retains all phase/band rows: nine high comparisons and 175 high initialization channels. All 15 electrical regression comparisons and 3,486 initialization channels remain high. Both scopes have zero skipped, missing, nonidentifiable, or deviating comparisons. The five-model origin attempt has three high comparisons, including RollingWheel, plus the existing GenerationOfFMUs EL005 refusal and a GyroscopicEffects Solve-phase timeout at 12.561 seconds. That timeout remains a failed normal-budget attempt. An isolated control with the phase watchdog absent takes 7.246 seconds in Solve lowering versus 7.423 seconds for the archived worker and produces identical compiler artifacts (bdf-gyro-solve-control-delta.json); it earns no normal-budget parity credit. Receipts are bdf-convergence-history-{canary,electrical,origin}-delta.json. The complete cohort comparison subsequently failed as recorded above. RollingWheel is still slower than OMC; no victory or release-ready claim is made.

Latest focused work: block-specific numerical AD seed domains

RollingWheel’s hottest projection JVP reads 47 seed coordinates, but only six belong to that block’s unknowns. The other 41 coordinates are fixed during the block solve. The compiler now issues a separate numerical AD kernel for the block’s exact seed domain, preserving ordered primal operations, whole-program outputs, canonical primal/JVP owners, and the unrestricted state/parameter derivatives. Partially active tensor loads split at domain boundaries; compact register-run packing preserves the exact ordered register values.

All 1,041 focused IR, AD, native-execution, and solver tests and affected-package all-target/all-feature Clippy pass (block-seed-domain-focused-3.log). Earlier attempts retain a missing re-export and a nesting lint. Regressions check fixed orientation matrix products, partial tensors of 4,096 elements, exact register bits, fresh Y/P/time, and stale canonical primal rejection.

The ordinary worker 4b4dc9c53ac993e1bc3a3d73d6d18d468bc26672305960213b5ae2cb06562dc8 takes 96.962 ms against the immediate archived 21e0c64e… recheck at 112.555 ms: 13.85% improvement for that pair. Flat, DAE, structural DAE, serialized Solve, and complete trace bytes are identical (block-seed-domain-profile-delta-1.json). The separate 40-run diagnostic has a 98.874 ms median (98.592–104.560 ms), with all complete traces bit-identical. The hottest selected program shrinks from 1,012 to 536 operations; its leaf samples fall from 561 in the earlier profile to 212 in the new profile. These are sampled profiles, not instruction timings. RollingWheel remains slower than OMC.

The first diagnostic attempt hit the 6 GiB worker limit because the temporary artifact probe duplicated shared general programs. Its exit 70 is retained and earns no validation credit. The bounded second probe dumps only specialized sources. All probe code is removed before ordinary verification; the restored worker source has SHA-256 0138785a9c71c923eb46793cbb95c5b069323b40af0f4df3fa0861244a24a5ee.

The fixed 20-model canary retains every phase and band: nine high comparisons, 175 high initialization channels, and zero skipped, missing, nonidentifiable, or deviating comparisons. The five-model origin check retains four high comparisons, 2,960 high initialization channels, and the existing GenerationOfFMUs EL005 refusal. Receipts are block-seed-domain-{canary,origin}-delta.json. The 15-model electrical regression set also retains every phase and band: all 15 comparisons and 3,486 initialization channels are high, with zero skipped, missing, nonidentifiable, or deviating comparisons (block-seed-domain-electrical-delta.json). These are focused checks, not a new cohort claim.

The subsequent complete multibody-block-seed-domain-full-11 sweep at clean commit 755c2169d66197da9d4853ff723a2017b9464658 passes with 158/566 strict-high models, including 22/42 MultiBody. All 158 comparisons and 20,379 initialization channels are high; the same 18 reviewed exclusions remain, with zero missing, nonidentifiable, or deviating comparisons. Every comparison band is unchanged. GearConstraint returns from the preceding standalone attempt’s timeout to its existing StructuralAnalysis EL005 refusal; it still produces no trace. Receipt: block-seed-domain-full-delta.json.

Latest focused work: prepare complete colored Jacobian applications

The candidate binds each projection’s colors and output placements to its immutable scalar-program owner. Scalar-program clones share their complete checked storage; canonical wire data and checked replay are unchanged. Native preparation retains code ownership and exclusive seed, output, and matrix scratch. Each call uses fresh coordinates, enters the external-table context once, preserves complete program order/status checks, and publishes only a successful matrix. Admission preserves mixed reverse/forward evaluation and the unrestricted initialization and parameter-sensitivity paths.

rolling-wheel/prepared-projection-profile-delta-1.json records 111.058 ms for the candidate worker f5c9fe7555cad3db316d2dca4061955091f9908401710e3ece3df4bfc0cc1ae7 and 126.144 ms for the immediately rechecked archived worker 0669bf79…. That pair improves 11.96%; it is not a repeated-median estimate. Flat, DAE, structural DAE, Solve, and complete trace bytes match. RollingWheel remains slower than OMC; no speedup over OMC is claimed.

The 897 focused tests and all-target/all-feature Clippy pass in prepared-projection-focused-3.log. Tests cover fresh coordinates, exact JVP placement, distinct source owners, retained native-code lifetime, external-table failure/recovery, mixed reverse admission, full parameter seed space, and atomic state preservation on failure. Earlier check logs retain a visibility error, a test dependency error, and a nesting lint; none counts as a passing run.

multibody-prepared-projection-canary preserves every phase and band: nine high comparisons and 175 high initialization channels, with no skipped, missing, nonidentifiable, or deviating comparison. The five-model multibody-prepared-projection-origin likewise preserves four high comparisons and 2,960 high initialization channels; the existing ThermalGenerationOfFMUs EL005 refusal remains. Receipts are prepared-projection-{canary,origin}-delta.json. The precursor passes the complete verify quick suite in verify-quick-prepared-projection-1.log: lint, 243 architecture tests, 17 file-size checks, 7,674 workspace tests, doctests, and all 28 pinned corpus models. Its complete MSL sweep preserves 158/566 strict-high models (22/42 MultiBody), all 20,379 initialization channels high, 18 tracked exclusions, zero missing or nonidentifiable traces, and zero deviations. Every phase and band matches the retained-workspace cohort (prepared-projection-quick-full-delta.json).

Review then found a stale-source counterexample: replacing the canonical JVP while retaining projection metadata silently used the old expressions. The reduced negative control fails in prepared-source-owner-red-1.log. Scalar views now retain the exact canonical owner when borrowing one scalar node; runtime rejects stale metadata and declines optional preparation for materialized views without that identity. The existing output-extent overflow check remains mandatory. All 1,090 focused tests and affected-package all-target/all-feature Clippy pass in prepared-source-owner-focused-2.log.

The final ordinary worker 21e0c64e50737cea054027ce6aee748fb02272e374f8cc13ab5087f603944885 runs in 113.944 ms against an immediate 134.107 ms archived-baseline recheck (15.03% for that pair). All five IR/trace artifacts remain byte-identical; prepared-source-owner-profile-delta-1.json records the comparison. The final multibody-prepared-source-owner-canary retains the same nine high comparisons, 175 high initialization channels, and all twenty phase/band rows, with zero skipped, missing, nonidentifiable, or deviating comparisons.

The final multibody-prepared-source-owner-full-11 sweep passes with 158/566 strict-high models (22/42 MultiBody), 158 comparisons, 18 tracked exclusions, zero missing/nonidentifiable traces, and zero deviations. All 20,379 initialization channels remain high. Every band matches the retained-workspace cohort; the existing GearConstraint failure changes from StructuralAnalysis EL005 to an attempt timeout, without producing a trace (prepared-source-owner-full-delta.json). The combined verify full attempt preserves every MSL phase and band, including GearConstraint’s existing EL005 refusal (prepared-source-owner-combined-full-msl-delta.json). Lint, architecture, 7,677 workspace tests, doctests, 28 pinned corpus models, examples, and binary builds pass. The template report test exposed a stale version-5 expectation against the canonical template schema version 6; the literal is corrected and all six report regressions pass (template-schema-version-6-2.log). Coverage increases from 77.27% to 78.34%, but the unchanged per-package uncovered-function ratchet fails. The failed gate report is retained beside verify-full-prepared-source-owner-1.log; no baseline is promoted. The attempt was stopped with exit 130 during documentation work after those failures to resume focused performance work. Remaining documentation/editor gates and a complete green release sweep are outstanding.

Verification checkpoint: remove asserted projection preconditions

rolling-wheel/verify-quick-projection-milestone-1.log preserves the failed combined quick run: MSL and the 28-model corpus passed, while test-helper lint, two added production assertions, and missing cargo-nextest blocked completion. The fixed tensor-update branch now consumes static-only subscripts. Logical output iteration borrows its checked source owner. Test helper refactoring retains every case and tolerance; the worker artifact test uses a normal success assertion. Neither architecture ceiling nor validation policy changes.

The focused checks pass: 316 Solve IR tests, all 243 architecture checks, 22 contact regressions, and all-target/all-feature Clippy for Solve IR, the worker, and the compiler facade. Logs are quick-blocker-focused-{1,2}.log. The first command stopped after architecture because its contact command named a nonexistent feature; the corrected command ran those tests successfully. The repository’s full Nix shell supplies nextest for subsequent combined verification. Combined quick/full are still pending, and this checkpoint makes no new performance or cohort claim.

Latest complete cohort: projection optimizations preserve all outcomes

target/msl/multibody-retained-sparse-workspace-full-11, at e4817572 with a clean tracked worktree, retains 158/566 strict-high models (27.92%), including 22/42 MultiBody models. All 158 compared models and 20,379 initialization channels remain high. The 18 reviewed exclusions remain; there are zero missing, nonidentifiable, or deviating comparisons. Every phase, simulation status, and comparison band matches multibody-native-inactive-tangents-full. This complete sweep covers grouped residual evaluation, streamed row scaling, and retained sparse factorization storage. Receipt: rolling-wheel/retained-sparse-workspace-full-delta.json.

The preceding multibody-retained-sparse-workspace-full attempt was interrupted with exit 130 while the initial memory budget restricted it to one worker. After the external memory-intensive job exited, the named -full-11 run used the requested eleven workers. The earlier incomplete log remains preserved and earns no credit; no model was retried within the completed sweep.

Removed timing probes identify the remaining cost. sim-step-probe-1 measures 1,054 accepted steps: 93.276 ms in numerical advances and 33.202 ms in the following host work. The latter includes 0.668 ms of sampler validation, 2.216 ms of event scanning, and 29.959 ms of endpoint/output work. Lease and initialization take 0.944 ms and plugin setup 0.465 ms. Its complete trace is byte-identical to the retained-workspace reference. sim-phase-probe-1 likewise shows initialization/setup around 1.5 ms, with identical trace bytes. These are attribution diagnostics, not additional coverage or comparative-speed measurements. Source archives and summaries are rolling-wheel/sim-{phase,step}-probe-{original,source}.json and rolling-wheel/sim-{phase,step}-probe-1-summary.json; all probes are removed.

RollingWheel remains slower than OMC. Combined quick/full verification and complete MultiBody coverage remain open.

Latest focused work: retain sparse numeric factorization storage

On top of c9e60096, the sparse projection owner retains faer’s numeric LU storage, exact conditioned coefficient bits, and a scratch buffer sufficient for both factorization and solving. The high-level API previously allocated a new numeric factor and scratch on each refactorization, then separate RHS, solution, and scratch buffers on each solve. The new owner uses the same faer symbolic/numeric kernels and pivot policy, with one owned result vector. Changed matrix patterns reconstruct the owner. Refactorization consumes the usable-factor state before modifying numeric storage; rejected factors cannot solve a RHS, and an identical rejected matrix is not refactored. Cloning owns independent factor and scratch storage. SPEC_0043 §6a states the ownership rule.

retained-sparse-workspace-focused-3.log records all 486 solver library tests and all-target/all-feature Clippy passing. Differential tests compare retained and freshly allocated faer solves bit for bit under coefficient, scaling, RHS, pivot, pattern, and dimension changes, plus singular failure/recovery and clone independence. Scratch addresses stay unchanged across repeated refactorizations. The first command had a test fixture integer-type error. The second caught an incorrect test assumption that every infinite coefficient produces a nonfinite solution; the final test compares faer’s existing behavior and checks recovery, without changing production acceptance or fallback policy.

The fixed target/msl/multibody-retained-sparse-workspace-canary retains nine high models and 175 high initialization channels. The five-model target/msl/multibody-retained-sparse-workspace-origin retains four high models and 2,960 high initialization channels. Both have unchanged phase/status/band outcomes, zero skipped/missing/nonidentifiable comparisons, and zero deviations. Receipts: rolling-wheel/retained-sparse-workspace-{canary,origin}-delta.json.

Controlled RollingWheel Sim is 0.125927 seconds versus the immediate archived baseline’s 0.136581 seconds, a single-pair 7.80% reduction. Flat, DAE, structural DAE, canonical Solve, and complete trace bytes are identical. Candidate worker SHA-256: 0669bf7946f1d278887c3940424f223b7ec3a81b6a6bf01da5c2c89f803c0916. Receipt: rolling-wheel/retained-sparse-workspace-profile-delta-1.json. The complete-cohort result above now covers this change. Combined quick/full verification remains open.

Previous focused work: stream structural row scales

On top of fce39f4e, sparse Jacobian row scaling consumes the pattern’s ordered borrowed row visitor instead of allocating every nonzero coordinate on each projection. Arithmetic, finite-contribution selection, and fallback scales are unchanged. The regression covers finite maxima and zero/nonfinite fallback rows against the dense implementation. Formatting, all 70 projection tests, and solver all-target/all-feature Clippy pass in streamed-row-scales-focused-1.log.

The fixed target/msl/multibody-streamed-row-scales-canary retains nine high models and all 175 high initialization channels, with unchanged phase/status/ band outcomes and no skipped, missing, nonidentifiable, or deviating comparisons. Receipt: rolling-wheel/streamed-row-scales-canary-delta.json. Controlled RollingWheel Sim is 0.133409 seconds versus the immediate archived baseline’s 0.136935 seconds, a single-pair 2.58% reduction. All four compiler IR artifacts and the complete trace are byte-identical. Candidate worker SHA-256: 754b3278ea9b2f246385c01e3143ec9665d0d0f81c09c34e74780b2a6f535978. Receipt: rolling-wheel/streamed-row-scales-profile-delta-1.json.

The attempted multibody-streamed-row-scales-full sweep was interrupted with exit 130 before completion. Its initial memory budget capped compilation and simulation at one worker despite the requested eleven. Its retained log is rolling-wheel/streamed-row-scales-full-1.log; it earns no cohort credit. The latest complete cohort below still predates both recent projection changes. RollingWheel remains slower than OMC; the next investigation is retained numeric factorization storage and scratch allocation.

The post-commit target/msl/electrical-grouped-residual-origin run at fce39f4e also retains all 15 high models and 3,486 high initialization channels, with unchanged phase/status/band outcomes and no skipped, missing, nonidentifiable, or deviating comparisons. Receipt: rolling-wheel/grouped-residual-electrical-delta.json.

Previous focused work: evaluate shared residual outputs once

On top of 02706ddf, construction binds each affine block’s selected residuals to unique, complete, repeatable canonical programs. The same output-selection utility now serves residuals and JVPs. Native and interpreted execution evaluate each selected program once and scatter its outputs in block order. Distinct owners stay distinct; assertions and failures remain part of complete program execution. Numerical projection, refinement, and acceptance are unchanged.

The RollingWheel dynamics block’s 24 residual outputs come from eight programs, three outputs apiece. The previous per-output API invokes each tensor program three times. grouped-residual-source-census.json records the source mapping. The longer repeated-sim-profile-1 diagnostic collects 6,062 leaf samples over 40 identical traces; 1,581 are in mapped JIT code. The main JVP program remains the largest kernel, and complete-input pure-call cache comparisons account for much of the sampled memcmp work. Repeated-run instrumentation is archived in repeated-sim-probe-source.json and removed from production source.

grouped-residual-red-2.log reproduces the absent grouped evaluation and failure propagation. The first RED command stopped at an unused-trait-method compiler error and is not a failing-test witness. Focused native tests prove one call per complete program, changing inputs, source output placement, prevalidation, explicit decline, and admitted table errors. Runtime tests cover changing coefficients and failure without replay or partial state commit. Ownership checks reject missing, ambiguous, and impure output selections. All 884 relevant library tests pass in grouped-residual-libraries-1.log; Clippy’s diagnostic nesting issues are corrected, with final affected-package Clippy green in grouped-residual-libraries-3.log and six grouped projection tests green in grouped-residual-libraries-2.log.

target/msl/multibody-grouped-residual-origin retains the same four high models and 2,960 high initialization channels. The fixed target/msl/multibody-grouped-residual-canary retains nine high models and 175 high initialization channels. Both have unchanged phase/status/band outcomes and zero skipped, missing, nonidentifiable, or deviating comparisons. Receipts: grouped-residual-origin-delta.json and grouped-residual-canary-delta.json. These are focused results; the complete cohort measurement below predates this change.

Controlled Sim is 0.144187 seconds versus an immediate archived baseline of 0.148633 seconds, a single-pair 2.99% reduction. Flat, DAE, structural DAE, canonical Solve, and the complete trace are byte-identical. Candidate worker SHA-256 is 4ca1831140c8a3ef38919a2e525c9469eb1f4988abac07864f95a96deddefc66. Receipt: grouped-residual-profile-delta-1.json. RollingWheel remains slower than OMC; no broader speed or complete-cohort claim is made.

Previous complete cohort: inactive AD tangents preserve the parity floor

target/msl/multibody-native-inactive-tangents-full, at c300c58d plus tracked worktree digest 0a5cc5ca811daf460bfc208afbce15ec01a29ae837671709aee525d4be01082b, retains 158/566 strict-high models (27.92%), including 22/42 MultiBody models. All 158 compared models and 20,379 initialization channels remain high. There are 18 reviewed exclusions, zero missing traces, zero nonidentifiable comparisons, and zero deviations. Every comparison band matches multibody-affine-conditioning-full. The already failing ChopperBuckBoost_DutyCycle reaches the checked interval-endpoint disagreement EX002 before its previous timeout; it remains a refusal and earns no credit. Concurrent RollingWheel Sim is 0.159451 seconds. Receipt: rolling-wheel/native-inactive-tangents-full-delta.json. Combined quick/full verification and complete MultiBody coverage remain open.

AD propagates constructor-proved zero tangents while retaining primal operations and their call owners, assertions, and domains. Solver-Y parameter loads may have zero tangents; full Y|P initialization and sensitivities retain parameter seeds. Inactive bilinear factors use the remaining product-rule term, with compact tensor packing and exact immutable-run transpose reuse. A numerically zero primal does not establish an inactive tangent. This follows SPEC_0039’s declared seed-domain rule and does not change Modelica equations.

native-inactive-tangents-focused-2.log records formatting, all 134 Solve library tests, and affected-package Clippy. Tests include finite-difference bilinear checks, parameter sensitivities, changing primal values, assertions, and retained invalid primal arithmetic. The first focused attempt failed formatting because restored sibling modules were absent; no tests ran until the complete candidate was restored. The fixed canary retains nine high models and 175 high initialization channels, with no status or band changes, skips, missing traces, or deviations. Receipt: native-inactive-tangents-canary-delta.json.

With bounded native SSA storage already landed, the general AD candidate measures 0.150045 seconds against the immediate archived baseline recheck of 0.152998 seconds. Both workers exit successfully and produce identical trace bytes (a81708c82927bf5721cf527c75b80f83c9c7fb97aedab05cda49da22d5e3c619). This single-pair 1.93% improvement is small; it is not a broad performance claim. Profiles are native-inactive-tangents-profile-1 and native-inactive-tangents-baseline-recheck-1; candidate worker SHA-256 is 68f4d57b6d43d1b86bd170a19f8c2dfc10f20fd8a31be8e33ed9f75ef4f149aa. The older pre-SSA trial remains rejected; its evidence is not reused as a pass.

A separate temporary diagnostic restricts each program’s Y seeds to the union of all projection unknowns selecting its outputs. It measures 0.137266 seconds with identical RollingWheel trace bytes (native-projection-seed-probe-1). The diagnostic changes a general derivative artifact and therefore is not a production implementation or coverage evidence. Its sources are archived in native-projection-seed-probe-trial.json and completely removed. A retained optimization needs a separate source-bound domain and must preserve unrestricted initialization and sensitivity derivatives. OMC’s generated system 731 solves six variables and reconstructs other forces/accelerations; Rumoca’s current certified affine dynamics block retains 24 variables. RollingWheel remains slower than OMC.

Previous complete cohort: fixed affine conditioning preserves all traces

target/msl/multibody-affine-conditioning-full, at b6a78cda plus tracked worktree digest 27b72247987f290dce47b2ab24944aeada03cbb440d69135beaf7451206a2b05, retains 158/566 strict-high models (27.92%), including 22/42 MultiBody models. All 158 compared models and 20,379 initialization channels remain high. Eighteen reviewed exclusions remain; there are zero missing, nonidentifiable, or deviating comparisons. Every phase, simulation status, and band matches multibody-inline-dae-view-full. Concurrent RollingWheel Sim is 0.163662 seconds, Driving 0.524226, and GyroscopicEffects 0.840462. Receipt: rolling-wheel/affine-conditioning-full-delta.json. Combined quick/full gates remain pending, and RollingWheel remains slower than OMC.

The temporary runtime-cost-probe-1 counters preserve the baseline trace byte for byte. Across that worker’s execution, the 24-unknown affine block has 2,794 Jacobian evaluations, 8,355 residual evaluations, and 5,561 linear solves. Its sparse cache performs 5,538 numeric factorizations: changing candidate-derived conditioning defeats reuse within the same fixed-matrix refinement. The counters include initialization checks, not only Sim. The active BDF session reports 1,054 accepted steps, 50 error-test failures, 2,245 nonlinear iterations, and zero nonlinear failures; its equation counters report 2,248 RHS calls and 160 Jacobian products, including setup probes. All temporary probes are archived in runtime-cost-probe.patch and removed.

Under the constructor-issued affine relation (SPEC_0043 §6), one invocation already retains its exact block Jacobian. Its conditioning now belongs to that same immutable matrix. Corrections solve with the original row/variable scales so the existing bitwise matrix cache can reuse LU. Acceptance still reevaluates the original residual, computes fresh candidate-derived scales, and performs the required small-coordinate correction. Nothing is reused across changed matrices by an approximate comparison. This follows the fixed-factor correction pattern in LAPACK DGERFS, while retaining Rumoca’s existing convergence policy and source residual.

The reduced offset-port electrical regression first fails the new conditioning reuse assertion (affine-conditioning-red-1.log). All 69 projection tests, including tiny junction voltages, opposing large currents, and fresh solution scales, then pass with affected-package Clippy in affine-conditioning-focused-2.log. The first focused command stopped at formatting; it did not run tests. The five-target origin retains four high MultiBody models and 2,960 initialization channels. The fixed canary retains nine high models and 175 initialization channels with unchanged statuses and bands. All fifteen models in the electrical affine-regression list compare high, with 3,486 initialization channels high and no skipped, missing, or deviating comparisons. Their bands match the corresponding members of the preceding full cohort; the older fifteen-model focused snapshot predates four already-landed recoveries, so its four apparent gains earn no new credit. Receipts: affine-conditioning-origin-delta.json, affine-conditioning-canary-delta.json, and affine-conditioning-electrical-delta.json.

The controlled candidate measures 0.153893 seconds against an immediate archived baseline recheck of 0.169451 seconds, a 9.18% reduction. Both exit successfully. Flat, DAE, structural DAE, and canonical Solve files are identical; floating-point traces differ and are validated by the OMC comparisons above. The candidate worker hash is 5b7273e9cc0b2e2464f810e6298d3cc4b71ebd6b8591fa1e38f9d1296f4cfa87. Receipt: affine-conditioning-profile-delta-1.json. The candidate’s 148 leaf samples include seventeen in JVP program 212 and five in sparse numeric LU; these self samples are distinct from invocation counts. Jacobian evaluation remains the next measured runtime target.

Prepared reverse-mode capability

PreparedScalarProgramBlock now derives its immutable per-program reverse-AD capability once, alongside existing row requirements, and preserves that metadata on clone. Previously every Jacobian call rescanned all operations, even to decline a multi-output tensor program. The capability predicate and input validation are unchanged; its cache is bound to the private immutable program block and contains no numeric results.

All 190 evaluator library tests, formatting, and Clippy pass in prepared-reverse-focused-1.log. The fixed twenty-member canary retains every phase, status, and band: nine compared models and 175 initialization channels high, with no skipped, missing, or deviating comparisons. Receipt: prepared-reverse-canary-delta.json. An isolated candidate measures 0.151358 seconds against an immediate 0.153747-second control. This single-pair 1.55% difference is small; it does not establish a broad runtime speedup. All Flat, DAE, structural DAE, Solve, and trace bytes match. Receipt: prepared-reverse-profile-delta-1.json. The complete-cohort claim remains bound to the preceding full run above.

Previous complete cohort: expression-view inlining restores the floor

target/msl/multibody-inline-dae-view-full, at fdef84cc plus tracked worktree digest 3f8fdd535362d6a4caf8c3279f4248ce82cf2dfb2f9bc65cad93bc70cbee50bd, restores 158/566 strict-high models (27.92%), including 22/42 MultiBody models. All 158 compared trajectories and 20,379 initialization channels are high; eighteen reviewed exclusions remain, with zero missing, nonidentifiable, or deviating comparisons. The complete band table matches multibody-identity-affinity-full. Driving completes Solve in 8.729543 seconds and GyroscopicEffects in 8.478000 seconds under the unchanged ten-second budget. The existing Media Inverse_sh_TX refusal changes ED019 to EF015; this earns no coverage. Receipt: rolling-wheel/inline-dae-view-full-delta.json.

Driving’s isolated Solve profile attributes 7.88% of self samples to ExpressionView::operation and 4.80% to DaeView::expression. These borrowed view accessors construct checked wrappers that consumers immediately inspect. Adding ordinary cross-crate inline hints exposes their bodies to consumer optimization without changing checks, data ownership, or evaluation. Isolated Solve falls from 8.574600 to 7.925615 seconds. DAE, structural DAE, canonical Solve, and trace files are byte-identical between direct-row-storage-driving-solve-profile-1 and inline-dae-view-driving-solve-profile-1.

Formatting, 43 evaluation-DAE tests, 187 DAE-IR tests, 125 Solve-phase tests, and affected-package Clippy pass in inline-dae-view-focused-1.log. The five-target origin retains four high models and all 2,960 initialization channels; the fixed twenty-target canary retains nine high models and all 175 initialization channels. Neither focused target set changes phase, simulation status, or band, and neither has missing or deviating comparisons. Receipts: inline-dae-view-origin-delta.json and inline-dae-view-canary-delta.json.

RollingWheel remains high with concurrent Sim 0.175470 seconds. It is still slower than OMC; no baseline promotion, complete MultiBody claim, or combined quick/full verification result is implied. Runtime profiling continues.

Previous complete cohort: native speedup exposed a Solve timeout

target/msl/multibody-direct-row-storage-full, at 6fd321b2 plus worktree digest 81a504337fd25db9d9e228bca672e474abd1a25f9b375805247b00dec8593c63, compares 157/566 models high (27.74%), including 21/42 MultiBody examples. All 19,487 compared initialization channels are high. Eighteen reviewed exclusions remain; no comparison is missing, nonidentifiable, or deviating. The prior 158-model floor is not preserved: RollingWheelSetDriving times out in Solve after 11.159 seconds against the unchanged ten-second budget. It completed Solve in 8.831501 seconds in the focused run. No timeout retry or budget change is used to recover coverage.

Fourbar_analytic and GearConstraint also reach their Solve watchdog instead of their previous structural refusals, and Media Inverse_sh_TX changes its existing frontend refusal back to EF015. None earns coverage. RollingWheel remains high with Sim 0.195300 seconds; the same run’s stored OMC simulation system timer is 0.133511 seconds. These concurrent timings are distinct from the isolated measurements below. The configured command exits zero, but the lost certified model prevents floor promotion. Receipt: rolling-wheel/direct-row-storage-full-delta.json. Combined quick/full verification and complete MultiBody coverage remain open.

Native performance repair: bounded tensor rows retain direct values

The seed-dependency census seed-frontier-1.log uses the production dependency deriver on every producer output of the checked RollingWheel JVP. Program 212 contains 1,106 producer operations: 571 wholly seed-independent, 128 mixed, and 407 wholly seed-dependent. All 31 matrix products and eight directional calls are mixed. Most independent operations are copies. Skipping only whole seed-independent operations would leave the expensive tensor work intact.

The native emitter exposes a separate concrete cost: any row-level tensor operation creates a register tape for every intermediate, including tiny three-vector operations. Existing direct lowering was already available for these operations. Under SPEC_0032 §§4–5 and SOLVE-C53, final native storage selection now retains direct SSA values when every tensor operation fits the existing 64-unit static-work bound. Matrix work includes contraction extent and AD lanes; other tensor ranges include their storage lanes. Larger tensors, overflowing work estimates, and tensor-update slices retain the prior tape selection. Canonical tensor owners and floating-point accumulation order do not change. The rule is independent of model identity or numerical values.

The reduced regression fails on the previous allocation policy in direct-row-storage-red-1.log. It then checks direct and tape-backed dual matrix products, two transposes, cancellation-sensitive accumulation, and changing seeds against the interpreter. A 4,096-element fill retains bounded loop storage and reloads changing inputs, including negative zero. All 88 native tests, formatting, and affected-package Clippy pass in direct-row-storage-focused-2.log; the first focused command stopped at formatting before running tests. All 99 compiler library tests, 595 core tests, seventeen architecture/spec gates, and 111 projection tests pass in direct-row-storage-core-1.log.

The controlled actual-worker trial measures Sim 0.168387 seconds. An immediate archived baseline recheck measures 0.287090 seconds: a 41.35% reduction. Both workers exit successfully. Flat, DAE, structural DAE, canonical Solve, and trace files are byte-identical; trace SHA-256 remains c9be85f47452aba0d450ab626c98a4cfdc838127d62df3eceea9edfd23f034e0. The hot residual kernel shrinks from 24,792 to 10,928 bytes and its JVP from 59,968 to 35,280 bytes. The trial worker is 599f608b7be9f51b262bf39fe452bb00ef58d4641deeff10650f512b80a6f7eb; its exact source patch and profile are retained in direct-row-storage-trial-1.patch and direct-row-storage-profile-1. The retained implementation moves the same selector to a dedicated module. Receipt: direct-row-storage-profile-delta-1.json.

The new profile has 168 leaf samples, 40 mapped to nonoverlapping JIT ranges. Sparse numeric LU contributes 21 samples, JVP program 212 contributes nineteen, and residual program 212 contributes four. These are self samples, not inclusive costs. Repeated algebraic factorization and JVP evaluation remain targets; this change does not establish that RollingWheel is faster than OMC.

The normal five-target origin comparison retains four high MultiBody models, all 2,960 initialization channels high, zero skipped/missing/excluded, nonidentifiable, or deviating comparisons, and the thermal GenerationOfFMUs EL005 refusal. RollingWheel Sim is 0.170735 seconds, GyroscopicEffects 1.018274, RollingWheelSetDriving 0.582798, and SphericalConstraint 0.177060. The fixed twenty canary members retain every phase, status, and band: nine models and all 175 initialization channels high, with eleven unchanged refusals and no unmeasured or deviating comparison. Receipts: direct-row-storage-origin-delta.json and direct-row-storage-canary-delta.json. The complete-cohort timeout above remains unresolved despite these focused checks.

Previous complete cohort: identity affinity preserves the restored floor

The complete target/msl/multibody-identity-affinity-full comparison at cc51f08ac23451e75ed9705085183900cc2e05ab retains 158/566 strict-high models (27.92%), including 22/42 MultiBody examples. All 20,379 compared initialization channels are high. Eighteen reviewed exclusions remain; there are zero missing, nonidentifiable, or deviating comparisons. The entire 566-row band table is unchanged from multibody-shared-materialization-full. Media Inverse_sh_TX changes its existing frontend refusal from EF015 to ED019; this earns no coverage credit and no model changes simulation status or band.

RollingWheel retains all 184 trajectory and initialization channels high. Concurrent Sim falls from 0.782661 to 0.327667 seconds; the controlled measurement below is 0.288191 seconds. It remains slower than OMC. Solve costs are SphericalConstraint 4.522983 seconds, RollingWheelSetDriving 9.211795, and GyroscopicEffects 9.539006. The latter two still need more margin under the unchanged ten-second budget. The run uses eleven simulation workers, the named commit’s tracked source, and the existing clean tracked-tree digest; the foreign untracked communication file remains untouched. Receipt: rolling-wheel/identity-affinity-full-delta.json.

All focused, fixed-canary, core, and architecture/spec checks recorded below pass. Combined quick/full verification and the remaining runtime performance gap remain open. This is not a baseline promotion, 100% MultiBody result, or release approval.

Runtime investigation: repeated coefficient evaluation

identity-affinity-symbols-profile-1 resolves the previous native-code blind spot without compiler instrumentation. Cranelift writes its JIT map when PERF_BUILDID_DIR is present at worker startup. The archived worker’s hash is 46722de1fb006b880281c8c9e7b213f7815cc9c4a0b65ce0c64fdc87570988af. Of 287 leaf samples, 148 map into the captured nonoverlapping JIT ranges. Program 212 accounts for 36 Jacobian samples and 35 residual samples, about 25% combined. These are sampled self costs, not inclusive call-chain costs. The three outputs still originate from Body.mo’s a_0 = der(v_0).

A private seed-domain probe binds each source program to the union of the unknowns of projection blocks selecting its outputs. Replacing other seeds with literal zeros alone changes Sim from 0.290190 to 0.293735 seconds and barely changes the hot Jacobian’s 59,968-byte code size. A subsequent prototype omitting structurally zero tangent calculations reduces the restricted kernel to 42,800 bytes and Sim to 0.274063 seconds. This is diagnostic evidence only: a restricted kernel cannot replace the general sensitivity artifact.

The attempted unrestricted production change also removes inactive bilinear terms and preserves compact storage through tensor fills, transposes, and concatenations. Its 134 Solve tests, 111 projection tests, 99 library tests, 595 core tests, seventeen architecture/spec gates, and Clippy pass. The five-target multibody-inactive-tangents-origin comparison retains four high models, all 2,960 initialization channels, zero skipped/missing/excluded or deviating comparisons, and the prior thermal EL005 refusal. However, the isolated candidate measures 0.304896 seconds, while an immediate archived baseline recheck measures 0.286091 seconds. Their trace files are identical. The extra packing outweighs the arithmetic savings, so the candidate and its temporary instrumentation are removed. Source, tests, and spec proposal are preserved privately in rolling-wheel/inactive-tangents-rejected-source.json; the normal worker was rebuilt and matches the archived baseline byte for byte. No new compiler capability or cohort improvement is claimed from this trial.

The next boundary is source coefficient evaluation across the entire solve. The read-only coefficient-frontier-2.log probe uses the production dependency deriver on every producer output of the current 24-unknown block. In program 212, 313 of 391 producer operations are independent of that block’s unknowns, including four of eight pure calls. Program 214 has 98 independent operations out of 112. Both programs occur in six of the block’s nine issued forward color selections; these are construction selections, not measured invocation counts. Merely reducing tangent arithmetic leaves this primal coefficient work inside each selected evaluation. OMC’s corresponding system 731 is a six-unknown linear solve with a symbolic Jacobian. The next experiment should bind a reusable coefficient evaluation to the exact immutable block coordinate and source owners, preserving assertions and fresh residual certification, then measure matrix assembly and residual evaluation together. General state/parameter sensitivities retain their complete seed domain. The latest complete cohort remains the 158/566, 22/42 checkpoint above; combined quick/full verification and the performance gap remain open.

Previous complete cohort: previous high-parity floor restored

The complete target/msl/multibody-shared-materialization-full comparison at f494d96207769c257168356e29fc5285b022fedb restores 158/566 high models (27.92%), including 22/42 MultiBody examples. All 20,379 compared initialization channels are high. The eighteen reviewed exclusions remain; there are zero missing, nonidentifiable, or deviating comparisons. Against the preceding multibody-fixed-anchor-full, GyroscopicEffects and RollingWheelSetDriving both recover their high bands. Against the earlier 158-model multibody-component-definition-full checkpoint, the complete 566-row band table is unchanged: no earlier high model has been exchanged for a different success. The run uses the unchanged phase budgets and eleven simulation workers. This restores the prior floor but does not establish 100% MultiBody coverage, a 28.00% floor, or release readiness.

GyroscopicEffects Solve measures 8.399072 seconds; RollingWheelSetDriving 9.541474; SphericalConstraint 4.684997. Driving remains close to the ten-second budget and still needs margin. Thermal HeatTransfer GenerationOfFMUs returns its explicit EL005 fixed-initial-value refusal without the earlier worker crash. Fourbar_analytic and GearConstraint reach structural refusals instead of timing out; Media Inverse_sh_TX changes its frontend refusal phase. These diagnostic changes earn no simulation coverage. RollingWheel retains all 184 trajectory and initialization channels high, with Sim 0.782661 seconds. It remains slower than OMC; the next runtime investigation is the repeated algebraic projection block and its causal assignment/tear selection.

Receipts are rolling-wheel/shared-materialization-full-delta.json and rolling-wheel/shared-materialization-full-restoration-delta.json. All tracked files match the named commit; the comparator’s dirty flag reflects the retained foreign untracked comm_fastdyn.md. Its tracked worktree digest is 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The focused and core validation below passes. Combined quick/full validation remains open; this cohort result is not a baseline promotion or release gate.

Focused runtime repair: identity tensors retain their constant degree

The current RollingWheel dynamics block has 24 unknowns and thirteen tear variables. Replaying its exact checked output dependencies with the existing candidate-aware tearing algorithm still produces thirteen tears, so changing the greedy candidate selection alone does not address the cost. Its issued affine certificate is false. OMC’s corresponding generated C uses a six-variable linear system (RollingWheel_03lsy.c, system 731, symbolic Jacobian method), rather than an analogous nonlinear solve.

The temporary affinity probe reads the captured checked Solve programs and locates the first divergence: 21 block rows prove affine, but rows 770–772, the three outputs of program 212, have no degree proof. Their source is Body.mo’s a_0 = der(v_0). Program 212’s second operation constructs an identity tensor; the degree checker lacks that constant operation and stops before reaching the remaining arithmetic. It has not proved a nonlinear equation. shared-materialization-tearing-inspect-2.log records the unchanged tearing counts, and shared-materialization-affinity-probe-1.log records the three missing proofs. Both diagnostics are read-only model replays.

The generic degree checker now recognizes a primal TensorIdentity as independent of all solver unknowns. The reduced regression fails in identity-affinity-red-1.log and passes after the change. It checks constant identity outputs and their affine product with unknown vectors at extents 3 and 4096, retaining compact ranges rather than enumerating matrix entries. The actual captured block then proves all 24 residuals affine (identity-affinity-probe-green-1.log). The temporary source probe is removed; all other unsupported operations, source evaluation, rank checks, and fresh affine-coordinate recovery remain unchanged under the existing SPEC_0036/0043 §6a degree-proof profile.

identity-affinity-focused-1.log passes all 315 Solve IR tests, 111 projection tests, formatting, and Solve IR all-target/all-feature Clippy. The normal target/msl/multibody-identity-affinity-origin five-target gate retains all four high MultiBody models, all 2,960 initialization channels, and the thermal GenerationOfFMUs EL005 refusal. No comparison is skipped, missing, excluded, nonidentifiable, or deviating. RollingWheel Sim falls from 0.782056 to 0.286439 seconds, retaining all 184 trajectory and initialization channels high. Receipt: rolling-wheel/identity-affinity-origin-delta.json, at 3723c597 plus worktree digest 17513b4fed4c226a5f93f65bed222c2b6f23a788242c188d88a7e0f8c2806f81.

The controlled actual-worker identity-affinity-rolling-profile-1 measures Sim 0.288191 seconds and 0.29 seconds user CPU, with zero major faults. The exact worker is archived for symbolization; 285 samples have zero reported loss. However, 49.47% of samples have unresolved shared-object identity and no JIT map was captured, so this profile does not yet attribute the remaining generated-code cost. It earns no separate coverage credit. The fixed twenty canary members retain their previous phases and bands: nine compared models and all 175 initialization channels high, eleven unchanged refusals, and no skipped, missing, excluded, nonidentifiable, or deviating comparison. Receipt: rolling-wheel/identity-affinity-canary-delta.json. RollingWheel remains slower than OMC; no closed-form trajectory or completed runtime optimization is claimed from this algebraic linearity proof. All 99 Rumoca library tests, 595 core tests, and seventeen architecture/spec gates pass in rolling-wheel/identity-affinity-core-suite-1.log. The complete cohort and combined quick/full verification have not yet been rerun after this change.

Previous complete cohort: fixed-anchor repair and remaining regressions

The complete target/msl/multibody-fixed-anchor-full run at cf0c2c61f01b914cfa76b1813c1ffaa4f293cc33 compares 156/566 models high (27.56%), including 20/42 MultiBody examples. The comparator measures 156 models, retains eighteen reviewed exclusions, and reports zero missing, nonidentifiable, or deviating traces. All 18,520 initialization channels are high. The configured CLI gate exits zero; the unresolved regressions still prevent promotion or release.

SphericalConstraint regains its high band, with Solve 5.225251 seconds and Sim 0.261808 seconds. GyroscopicEffects again exceeds the unchanged ten-second Solve budget, despite passing the preceding focused comparison in 9.499843 seconds. RollingWheelSetDriving loses its prior high band to the same Solve budget; its preceding full-run Solve time was already 9.430834 seconds. These failures remain measured failures, without retries or changed limits. The elementary RollingWheel retains all 184 trajectory and initialization channels high; Sim is 0.830653 seconds, still slower than OMC.

The full run also changes Thermal HeatTransfer GenerationOfFMUs from its previous typed fixed-initial-value refusal to a worker stack overflow (SIGABRT). This is an actionable compiler regression to reduce and repair, even though the model had no prior high band. The original log records rumoca-worker-main stack exhaustion. PrismaticConstraint now reaches a structural refusal instead of timing out, and Media Inverse_sh_TX changes its frontend refusal phase; neither earns coverage credit. The complete per-model receipt is rolling-wheel/fixed-anchor-full-delta.json. The next work is the stack-overflow root cause and remaining Solve costs; no full-cohort restoration is claimed from the focused three-model success. The subsequent focused repair below addresses the crash without revising this cohort result.

Focused repair: cancellation retains a shaped primal zero

The actual-worker rolling-wheel/fixed-anchor-thermal-profile-1 reproduces the thermal GenerationOfFMUs stack overflow. Its core dump repeatedly enters tensor_coefficient while attempting to recover the shape of a zero primal. The temporary source-DAE probe in cancelled-offset-source-probe-1.log locates the exact recipe: two references to the same shared zero offset are subtracted and negated. The source operand is the heat capacitor’s der_T coordinate. This is not a cycle in the source expression DAG. A separate source-level theta-theta control already passes; the failure is in reconstruction of the derived shared coefficient recipe.

The prior derivative repair reused exact-sum differentiation for order-zero values. Its Derivative::Zero result lost the primal shape, and shape recovery requested the same cancelled primal recursively. Primal cancellation now constructs a zero with the identical checked operands’ shape before returning or caching its value. Positive derivative orders retain the structural-zero tag. Reconstruction uses the existing shared shaped_zero constructor, with no duplicated zero builder, tensor scalarization, raised stack limit, or model-specific condition. The original equations and initialization checks remain authoritative under SPEC_0007/0040 STRUCT-T03 and MLS §8.6.

The reduced checked-DAE regression reproduces the captured recipe and checks scalar, vector, matrix, and empty-tensor values through derivative order two with the independent DAE numeric evaluator. cancelled-offset-shape-red-1.log records the failed shape-recovery boundary; cancelled-offset-focused-2.log passes all 166 structural tests, 22 contact tests, formatting, and structural all-target/all-feature Clippy. Temporary diagnostics have been removed.

The normal target/msl/multibody-cancelled-offset-origin gate compares all four selected MultiBody models high, with all 2,960 initialization channels high and zero skipped, missing, excluded, nonidentifiable, or deviating traces. The fifth target, thermal GenerationOfFMUs, again returns its prior explicit EL005 fixed-initial-value refusal for inverseCapacity.mass.T; the crash is repaired, but this is not thermal-model simulation support. SphericalConstraint Solve is 4.906882 seconds; GyroscopicEffects is 9.893319 seconds; and RollingWheelSetDriving is 9.252459 seconds. These near-budget Solve costs still require profiling. RollingWheel Sim is 0.784325 seconds and remains slower than OMC. Receipt: rolling-wheel/cancelled-offset-origin-receipt.json, at 9cd80190 plus worktree digest 2cff66a4996d90fafb05aec84234a5594c0a7d5dad05c5d85c8c9a18a22bbebe.

The same-source target/msl/multibody-cancelled-offset-canary retains all twenty phase, status, and band rows from multibody-fixed-anchor-canary: nine compared models and 175 initialization channels high, eleven unchanged refusals, and zero skipped, missing, excluded, nonidentifiable, or deviating comparisons. Receipt: rolling-wheel/cancelled-offset-canary-delta.json. All 99 Rumoca library tests, 595 core tests, and seventeen architecture/spec gates also pass in rolling-wheel/cancelled-offset-core-suite-1.log. The complete cohort and combined quick/full suites have not been rerun after this repair; the remaining near-budget MultiBody Solve costs are still open.

Focused performance repair: reuse complete source dependency witnesses

The actual GyroscopicEffects worker spends 8.534372 seconds in Solve at 79347dca, with 8.48 seconds of user CPU. The recorded Solve profile includes repeated source materialization walks, expression decoding, and allocations. Source inspection shows that each affine candidate starts the same root dependency proofs again, allocating a whole-DAE visited array per query. An initial singleton-projection prefilter avoids constructing a numeric evaluator for ineligible aggregate shapes and literal integer indices, but alone only changes the controlled Solve measurement to 8.424262 seconds.

Affine discovery now owns a cache bound to one immutable DAE view and facts pass. Its key retains the exact source expression and function call path; its value retains the complete state-anchor witness or the original refusal. Every candidate still checks its own state against that witness. No cached boolean independence result or visited marker can erase an anchor, and no cache survives a facts-update boundary. This preserves the STRUCT-T03 independence contract and source tensor ownership. New checked-DAE tests exercise alternating candidate states, nonlinear self-dependence, shared function bodies with distinct arguments/refusals, and extents 3 and 4096.

shared-materialization-focused-1.log passes 168 structural and 22 contact tests; its test-only nesting lint was corrected, and shared-materialization-focused-2.log passes the two new tests, formatting, and all-target/all-feature structural Clippy. The normal five-target target/msl/multibody-shared-materialization-origin comparison has unchanged phases and bands against multibody-singleton-prefilter-origin: four models and all 2,960 initialization channels high, with zero skipped, missing, excluded, nonidentifiable, or deviating comparisons. Thermal GenerationOfFMUs retains its prior EL005 fixed-initial-value refusal without a worker crash. GyroscopicEffects Solve measures 8.967672 seconds; SphericalConstraint 4.596436; RollingWheelSetDriving 9.132433. RollingWheel retains its 184 high channels and Sim measures 0.782056 seconds, still slower than OMC. Receipt: rolling-wheel/shared-materialization-origin-delta.json, at 79347dca plus worktree digest bfd38449f5f27a9b11501de6973429b0e48a98f1d56404a28f9a4977cda696d9.

The controlled actual-worker shared-materialization-gyro-solve-profile-1 measures Solve 7.759928 seconds and 7.72 seconds user CPU, versus the 8.534372 baseline. The profile has zero lost samples and archives its exact worker binary for subsequent symbolization. This single-worker diagnostic bypasses the parent Solve watchdog and earns no coverage credit; the normal gate above supplies the parity evidence. The fixed twenty-model target/msl/multibody-shared-materialization-canary has unchanged phases and bands against multibody-cancelled-offset-canary: nine compared models and 175 initialization channels high, eleven unchanged refusals, and zero skipped, missing, excluded, nonidentifiable, or deviating comparisons. Receipt: rolling-wheel/shared-materialization-canary-delta.json. All 99 Rumoca library tests, 595 core tests, and seventeen architecture/spec gates pass in rolling-wheel/shared-materialization-core-suite-1.log. Full-cohort restoration is not established by these focused measurements; combined quick/full verification remains open.

Previous complete cohort: two state-reduction regressions recorded

The complete target/msl/multibody-no-slip-full run at 5be301f43b2738cde950733e59de1e5e57aa9faa compares 156/566 models high (27.56%), including 20/42 MultiBody examples. Eighteen reviewed exclusions remain; there are zero missing, nonidentifiable, or deviating comparisons. All 18,495 compared initialization channels are high. Although the configured CLI gate exits zero, this is a regression against the preceding 158-model checkpoint and is not acceptable for promotion or release.

SphericalConstraint and GyroscopicEffects both lose their prior high bands because Solve exceeds its unchanged ten-second parent budget. The preceding Solve measurements were 5.220201 and 8.457811 seconds, respectively. Their failures remain visible in rolling-wheel/no-slip-full-delta.json; no timeout or comparator policy was relaxed. RollingWheel retains all 184 trajectory and initialization channels high with eight states, and Sim measures 0.835341 seconds in this cohort run. This still does not meet the OMC speed objective.

The separate no-slip-spherical-solve-profile-1 diagnostic reveals that SphericalConstraint’s expensive reduction attempts end in a structural refusal: 2194 matched of 2219 equations and unknowns. This diagnostic invokes the worker directly and does not apply the cohort parent’s ten-second Solve watchdog; it does not replace the timed-out cohort result or earn coverage credit. Profiling was enabled at Solve entry and retained through the refusal. Leading self costs include DAE operation/view access, projected incidence traversal, and materialization proofs. The controlled reproduction below locates the responsible state-reconstruction decision.

The source-DAE inspection in no-slip-spherical-reduction-2.log selects a new auxiliary definition for freeMotionScalarInit.initAngle.angle after three direct state demotions, and later selects one for constraint.frame_b.r_0. The direct-first lane stalls after reducing the unmatched residue from fifty to six; the pristine holonomic-first attempt also fails. The initial hypothesis that the angle identity solve hides explicit component kinematics is rejected by the controlled comparison below.

Fixed-anchor regression: retaining exact derivative zeros

The source-DAE controls in rolling-wheel/no-slip-spherical-aux-control-2.log isolate constraint.frame_b.r_0: disabling only its new auxiliary definition restores structural reduction in 3.885 seconds. Disabling only the angle definition still fails after 21.757 seconds. These temporary diagnostic controls have been removed from the compiler and retained only in the private campaign evidence. The production fix makes no decisions from model names or variable ordinals.

The reduced TensorFixedAnchor.mo fixture contains eighteen scalar equations and unknowns. Its constant matrix constrains a position vector to a fixed anchor, while source derivative equations define velocity and acceleration. Before the fix, structural reduction matches only twelve of eighteen unknowns. OMC eliminates all dynamic states and agrees with the independent analytical solution on all eighteen channels over twelve rows, with zero observed error (rolling-wheel/omc/fixed-anchor-1/analytical-comparison.json).

The first divergent layer is auxiliary tensor differentiation. For A*q=b, it materialized a zero derivative matrix and retained A'*q even when A'=0, creating a false primal-coordinate dependency. Coefficient reconstruction and its shared cache now preserve the existing Derivative::Zero distinction. Only nonzero coefficient derivatives read the primal coordinate or its first derivative. The original checked aggregate solve with A remains authoritative, including its nonsingularity requirement. This follows MLS array equality and continuous-equation semantics, SPEC_0022 §3.7, and SPEC_0007/0040 STRUCT-T03. It introduces no scalarized tensor owner.

The reduced fixture now has no dynamic states and checks all eighteen analytical channels with both BDF and RK. A singular constant-matrix variant is still rejected by both integrators. All 165 structural library tests, 22 contact tests, formatting, and structural all-target/all-feature Clippy pass in rolling-wheel/fixed-anchor-focused-2.log.

The normal target/msl/multibody-fixed-anchor-origin comparison measures all three selected models high, with zero skipped, missing, excluded, nonidentifiable, or deviating traces. All 2,068 initialization channels are high. SphericalConstraint completes Solve in 4.898042 seconds and Sim in 0.247828 seconds with thirteen states; GyroscopicEffects completes Solve in 9.499843 seconds and Sim in 1.494477 seconds with eighteen states. The original ten-second Solve budget is unchanged, leaving GyroscopicEffects little margin. RollingWheel retains eight states and all 184 channels high, with Solve 1.269984 seconds and Sim 0.782680 seconds. It remains slower than OMC. The receipt is rolling-wheel/fixed-anchor-origin-receipt.json, at 7602af65 plus worktree digest 786ae41e9185c7215630f8c3892cef53bf52aa7f8fd07eb20773e946c93ef3e2.

The same-source fixed twenty-model canary target/msl/multibody-fixed-anchor-canary has no phase, status, or band changes against multibody-no-slip-canary-2: nine compared models and 175 initialization channels remain high, with eleven unchanged refusals and zero skipped, missing, excluded, nonidentifiable, or deviating comparisons. Receipt: rolling-wheel/fixed-anchor-canary-delta.json. Complete-cohort validation remains pending; these focused checks do not revise the measured cohort above. The broader focused checkpoint passes all 99 Rumoca library tests, 595 core tests, and seventeen architecture/spec gates in rolling-wheel/fixed-anchor-core-suite-1.log. The combined verify quick and verify full suites have not been rerun for this checkpoint.

Previous complete cohort: tensor-state reconstruction preserves parity

The complete target/msl/multibody-component-definition-full gate passes at 27dd45f57af983cae6faf5f6ec9c2b7936129f4c, using eleven Rumoca workers, the unchanged 566-model roster, and unchanged tolerances and budgets. All 158 previously strict-high models retain their bands: 158/566 (27.92%), including 22/42 MultiBody examples. The comparator measures 158 models, retains eighteen reviewed exclusions, and reports zero missing traces, nonidentifiable traces, or deviating channels. All 20,379 initialization channels are high.

RollingWheel completes the normal pipeline with eleven integrated coordinates and all 184 trajectory and initialization channels high. Sim measures 1.071142 seconds. This establishes cohort safety for the tensor-state change; it does not establish a meaningful speed improvement or erase the preceding single-model startup failure. The only phase/status delta is the already excluded ThyristorBridge2Pulse_RLV_Characteristic, which completes instead of timing out and returns the exclusion count from seventeen to eighteen. No comparison band changes and no coverage credit comes from this timing delta.

Receipt: rolling-wheel/component-definition-full-delta.json; worktree digest: 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. OMC version is a96aa1a-cmake. RollingWheel remains slower than OMC; its velocity reconstruction and the combined quick/full gates remain open.

Latest focused work: reconstructing dependent velocity tensors

TensorNoSlipContact reproduces the remaining RollingWheel velocity-state problem. Its contact geometry and two no-slip equations determine the velocity vector from independent position and angular coordinates. OMC selects three states (theta, x, y); Rumoca initially retains the three additional velocity coordinates. Independent analytical motion is theta=0.2+t, x=0.2, y=0.3+sin(0.2)-sin(theta), and z=cos(theta), with velocity and force given by their first and second derivatives. All 21 OMC channels agree with that solution to 1.079e-8 over twelve rows in rolling-wheel/omc/no-slip-reconstruction-1/analytical-comparison.json.

The STRUCT-T03 auxiliary-state profile now combines independent source scalar constraints with independently materializable entries of authored literal-array equalities. Coefficients remain aggregate identity, projection, transpose, product, and outer-product expressions; there is no tensor-basis enumeration. The original equations, initialization obligations, and assertions remain owners. Self-dependent coefficients, incomplete systems, and causal projections that merely repeat their own definitions are refused. A direct source state definition takes precedence over a redundant auxiliary solve for that state.

Tracing the real wheel immediately after its four direct state demotions found a further discrepancy: the equality proof knew that the joint velocity equaled the body’s velocity state, but affine reconstruction required a unique causal assignment. The joint has multiple source defining equations, so that lookup returned none. The affine proof now consumes the existing signed, offset-free state-alias proof. It recognizes both no-slip rows and the independent vertical velocity row before holonomic differentiation. Position-array discovery and projection follow the same signed value aliases. Positive and negated velocity aliases have reduced analytical regressions; signed literal arrays retain their signs when contributing an independent row.

The normal target/msl/multibody-no-slip-alias-origin gate succeeds with eight integrated coordinates (x, y, three angles, three angular rates), matching OMC’s state set, and no retained manifold rows in the structural diagnostic. All 184 trajectory and initialization channels are high: one model compared, zero skipped, missing, excluded, or nonidentifiable traces. Sim is 0.801544 seconds versus 1.068513 seconds in the preceding eleven-state focused run; Solve is 1.284566 seconds and backend build 1.165971 seconds. This remains much slower than OMC and establishes no new cohort coverage. The focused receipt is rolling-wheel/no-slip-alias-origin-delta.json, with worktree digest f8e606b2f16621b50aaa0e2cb8107524c2b3f02d6e02447a2583e32e645199d0.

All 165 structural library tests, nineteen contact tests, formatting, and structural all-target/all-feature Clippy pass in no-slip-focused-suite-6.log. The additional negated-position case also passes within the complete 593-test core suite; all 99 compiler library tests and seventeen gates pass in no-slip-core-suite-1.log. Temporary source-DAE probes have been removed; their results remain in no-slip-intermediate-{4,alias-5}.log and no-slip-real-model-alias-6.log.

The fixed multibody-no-slip-canary-2 passes with all twenty phase, status, and band rows unchanged from multibody-component-definition-canary: nine models compared high, 175 initialization channels high, eleven unchanged refusals, and zero skipped, missing, excluded, nonidentifiable, or deviating comparisons. Receipt: rolling-wheel/no-slip-canary-2-delta.json; worktree digest: 3c24b570f8d3970f81bc080d6a5b74bc0e09134f2788ab626ab2cef6e90dfe35. The first multibody-no-slip-canary was externally interrupted during library parsing and has no parity measurement; its log remains preserved. No solver failure was retried, and no budget or tolerance changed.

The controlled actual-worker no-slip-profile-1 measures Sim at 0.795771 seconds (0.78 seconds user CPU, zero major faults during Sim). Every native sample address maps to a generated function. Projection Jacobians account for 13.23% of sampled self time, assignment schedules 10.51%, and residual programs 7.54%; there are no manifold Jacobian programs. Source-bound programs 212 and 214 originate at Parts/Body.mo’s a_0=der(v_0) and z_a=der(w_a) equations. Their projection Jacobians alone account for 8.55% and 3.38%, respectively. Runtime interpretation, assignment certification, and allocation remain visible costs. The worker SHA-256 is 4ef449ffd30e161cf8030f8eaf130f568a3ebb1515fb7cda6ccdfa2fc5e7939f. This is a performance diagnostic, not separate parity or cohort evidence. The next step is a complete cohort at this implementation, followed by tracing those acceleration kernels and prepared projection execution against OMC.

Previous focused work: inverse coordinates of tensor states

TensorStateContact reproduces RollingWheel’s source pattern: a position state array equals {x,y,z}, its derivative is a velocity array, and independent orientation constraints determine the contact vector. Structural discovery could not derive z=-delta[3] through those array equalities. The RED test keeps five state declarations (nine scalar coordinates), while OMC selects only theta, x, and y. Its generated equations and eighteen analytical channels agree to 4.44e-16 over twelve output rows in rolling-wheel/omc/component-definition-1.

The STRUCT-T03 component profile now follows unique literal-array definitions of non-Always states while discovering their inverse scalar coordinates. It captures one component expression for proof and reconstruction, rejects cyclic/self-dependent definitions, and retains the source tensor owners, assertions, and initialization obligations. Both ordinary residuals and the existing row-major structured bodies participate. The reduced model reaches three states under BDF and RK, with independently checked position, velocity, and force. Reversed and zero-residual forms agree; source assertions still execute; a free-fall negative control retains its free position and velocity. All fourteen contact tests, 162 structural library tests, 99 compiler library tests, 587 core tests, seventeen gates, formatting, and structural Clippy pass.

An initial broader component-definition candidate made six existing contact tests structurally singular by introducing unrelated scalar definitions into holonomic discovery. The final profile admits only inverse coordinates of the explicit state-array definition. All affected tests pass again. The initial free-fall fixture also needed two missing free horizontal variables to balance its intended equations. RED/intermediate evidence remains in rolling-wheel/component-definition-{red-1,probe-1,probe-2,contact-1}.log.

The normal multibody-component-definition-origin attempt failed its unchanged sixty-second worker startup budget during source loading, before model compilation; its parity is unmeasured. A separate archived-DAE diagnostic and the controlled actual-worker component-definition-profile-1 both complete and produce byte-identical eleven-state traces with SHA-256 91cd3b96b253f0b05162a29f22564644ef595acece2884f844a37dffafed7184. The position array is reconstructed; the three body velocity coordinates are still integrated. Existing plot-compare --reuse-traces compares all 184 trajectory channels against OMC: the maximum bounded channel score is 1.188e-4, below the unchanged high threshold. These diagnostics do not replace the failed normal attempt or provide cohort coverage credit.

The controlled profile measures Sim at 1.062014 seconds (1.05 seconds user CPU). This does not establish a meaningful speed gain over the preceding 1.088924 seconds, and remains slower than OMC. Projection Jacobian programs consume 15.60% of sampled CPU self time, manifold Jacobians 12.43%, and typed directional calls 11.82%; all JIT sample addresses map to generated code. Worker SHA-256: bca1176e41b141f86e3d49c2cd9f56b15ce9128b24012b2f7286bf67351ef5fc. The next unresolved proof is reconstruction of the remaining velocity vector from the no-slip constraints and independent angular motion. The original scalar-state TensorContact still retains three states.

The fixed multibody-component-definition-canary retains every phase, status, and band from multibody-direct-auxiliary-canary: nine models compared high, 175 initialization channels high, eleven unchanged refusals, and zero skipped, missing, nonidentifiable, or deviating results. Receipt: rolling-wheel/component-definition-canary-delta.json; worktree digest: b98745ba9ee53b6a18d5d99a9ef2b3501a67c8cc9922b7f0b9fd9995f15edbc1. The complete cohort above validates this component reconstruction change.

Previous complete cohort: auxiliary derivative reconstruction

The complete target/msl/multibody-direct-auxiliary-full gate passes at ae6dbf5bab8be88da2b203ad38698ab98ef32adf, with eleven Rumoca workers, unchanged budgets and tolerances, and the same 566-model roster. It measures 158/566 strict-high models (27.92%), including 22/42 MultiBody examples. All 158 compared models and all 20,379 initialization channels are high; seventeen reviewed exclusions remain, with zero missing, nonidentifiable, or deviating traces. Every previously strict-high model retains its band.

GyroscopicEffects completes Solve in 9.963 seconds and simulation in 1.506 seconds, gaining a high comparison where the preceding run timed out in Solve. This is close to the unchanged ten-second Solve budget; it does not establish a robust performance margin or attribute the gain to this structural change. IMC_Transformer reaches its existing structural refusal instead of timing out. The already excluded ThyristorBridge2Pulse_RLV_Characteristic times out during simulation instead of completing: its failure remains visible and accounts for the exclusion count dropping from eighteen to seventeen, without changing a comparison band. No coverage claim relies on a retry.

Receipt: rolling-wheel/direct-auxiliary-full-delta.json; worktree digest: 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. OMC version is a96aa1a-cmake. RollingWheel’s OMC speed target and combined quick/full remain outstanding.

Previous focused work: preserve auxiliary derivatives across reduction rounds

The explicit-height control of TensorContact retained three states despite the existing auxiliary proof determining its contact vector from orientation. Direct-state derivative preflight did not consume that proof. Once admitted, height demotion succeeded, but velocity demotion stopped at a generated LinearSolve expression. Supporting its derivative then exposed a Solve ownership rejection: a bare aggregate solve cannot enter scalar lowering. The retained RED and intermediate failures are rolling-wheel/direct-auxiliary-{red-1,candidate-1,candidate-2}.log.

Preflight now follows the auxiliary block’s exact operand contexts and excludes the state being demoted. First and second derivatives of a checked linear solve use the implicit matrix equations on the same nonsingular domain. Reconstructed solves retain aggregate function owners, generated once per checked Real matrix/vector signature, with distinct source-bound invocation identities. SPEC_0007’s STRUCT-T03 profile and its SPEC_0040 reconstruction rows govern this extension; the existing supplied-derivative rule moved into those referenced rows to keep the parent within its word budget.

The explicit-height control now uses one state and preserves all analytical channels under both BDF and RK. The reciprocal-motion regression verifies q''=2/(1+theta)^3 when theta'=1: the matrix and right-hand side have zero second derivatives, so the mixed term is essential. AngularRateContact also drops from three state declarations to two while preserving its analytical motion. All eleven contact tests, 162 structural tests, 99 compiler library tests, 584 core tests, 17 gates, formatting, and structural-package Clippy pass. The reciprocal fixture’s first attempts needed the checked optional fixed attribute and an unfixed algebraic start guess; these were fixture corrections, not compiler failures or additional coverage.

The normal target/msl/multibody-direct-auxiliary-origin gate compares one RollingWheel trace: all 184 trajectory and initialization channels high, zero skipped/missing/nonidentifiable/deviating results. Sim measures 1.055705 seconds, with the same fourteen integrated coordinates and byte-identical trace SHA-256 f46cbcf4bf00620007139eb17416c2ece8c3ddf5665e577a3c4e22ef2dadc47c. This does not establish a meaningful RollingWheel speed improvement. The controlled direct-auxiliary-profile-1 attempt expired after 180 seconds in source loading, before Sim; it provides no profiling measurement.

The fixed multibody-direct-auxiliary-canary retains every phase, status, and band from multibody-nested-call-sharing-canary: nine models compared high, 175 initialization channels high, eleven unchanged refusals, and zero skipped/missing/nonidentifiable/deviating results. Receipt: rolling-wheel/direct-auxiliary-canary-delta.json; worktree digest: 90691544006cba292c109a5c2d88f90f894836c0920f6d0eb7d1f3ae706ec964. The original tensor-form contact model still retains three states. Its missing component-definition proof, RollingWheel’s extra state coordinates, the OMC speed target, and combined quick/full remain open. The complete cohort above validates this focused structural change.

Previous complete cohort: execution optimizations preserve parity

The complete target/msl/multibody-nested-call-sharing-full gate passes at eb7f32ec05bc98eed453a5ad643f9b777d968e44, using eleven Rumoca workers and the unchanged budgets, tolerances, and 566-model roster. It confirms the three focused optimizations below: 157/566 models remain strict-high (27.74%), including 21/42 MultiBody examples. The comparator measures 157 models, retains eighteen reviewed exclusions, and reports zero missing traces, nonidentifiable traces, or deviating channels. All 19,412 initialization channels are high. Every model retains its preceding comparison band.

Seven unsuccessful models change phase/status: the six Fluid startup failures from multibody-causal-ready-full reach their existing typed frontend refusals, while IMC_Transformer times out instead of reaching its prior EL005 refusal. These are recorded as failures, with no new semantic coverage credit. Receipt: rolling-wheel/nested-call-sharing-full-delta.json; worktree digest: 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. OMC version is a96aa1a-cmake. The OMC speed target and combined quick/full remain outstanding.

Previous focused work: share exact nested source call owners

On top of 0b273fb17391f0ce960e718a4ce20cd1d0670884, pure-call registration now reuses completed nested DAE call owners within the same Solve table. Previously only root calls shared this lookup, expanding a shared source call graph into an invocation tree. The reduced five-owner example produced seven owners before the repair. Distinct source calls remain distinct even with identical names, spans, or bodies; the active-function recursion check still precedes reuse. Actual arguments, directional seeds, and each invocation’s assertion predicates retain their own execution coordinates. Finishing a table discards the registration scope. This refines the existing pure-call construction contract in SPEC_0043 under SOLVE-C51/C52.

All three focused regressions pass, covering increasing graph depth, different arguments/seeds and predicate outcomes, and table-scope reset. All 315 Solve-phase/evaluator library tests, 99 compiler library tests, 582 core tests, 17 gates, formatting, and affected-package Clippy pass. RED/GREEN evidence is in rolling-wheel/nested-call-sharing-{red-1,green-2}.log.

RollingWheel’s canonical call table shrinks from 1,216 to 214 owners. In the controlled actual-worker profile nested-call-sharing-profile-1, native build time falls from 2.476045721 to 1.516248235 seconds. Sim measures 1.088924 seconds (1.07 seconds user CPU), versus 1.134402 in the preceding controlled profile; the trace remains byte-identical with SHA-256 f46cbcf4bf00620007139eb17416c2ece8c3ddf5665e577a3c4e22ef2dadc47c. Worker SHA-256 is c8fc7489fda11426c4f43c3d228e3cc08866bc2c7c737f22e1f6e126ac98da6f. Solve construction itself does not improve in this measurement. Native algebraic projection and manifold Jacobian programs still account for 14.30% and 11.91% of CPU samples, respectively; these are self costs.

The normal target/msl/multibody-nested-call-sharing-origin gate measures Sim at 1.073520 seconds and compares one model: all 184 trajectory and initialization channels high, with zero skipped, missing, nonidentifiable, or deviating results. The fixed multibody-nested-call-sharing-canary retains every phase, status, and band from multibody-typed-small-tensor-canary: nine compared models high, 175 initialization channels high, eleven unchanged refusals, and zero skipped/missing/nonidentifiable/deviating results. Receipt: rolling-wheel/nested-call-sharing-canary-delta.json; worktree digest: d928c2253b88de30dbe8b40378311c89483c6bafcba6943c944aeec04e686481. The complete cohort above confirms these three execution/construction optimizations. The OMC speed target and combined quick/full remain outstanding.

Previous focused work: emit small native tensor arithmetic directly

On top of 6306f702d820717ad506fe705e0bb35253b43158, typed native function helpers now use the same bounded direct matrix arithmetic as scalar kernels. Previously even a 3-by-3 product emitted two counted loops and dynamic tape addresses. Products costing at most 64 terms now emit ordered multiply/add operations; short elementwise operations emit at most sixteen elements directly. Larger operations retain loops. This is final backend emission under SPEC_0032 §4 and SOLVE-C53: canonical tensor owners, call identities, and floating-point accumulation order stay unchanged.

An isolated million-call 3-by-3 diagnostic improves from 86.9 to 51.8 ms; the temporary throughput probe is removed. The two permanent regressions cover vector/matrix orientations, both sides of the emission-size boundary, cancellation-sensitive accumulation, and changing directional seeds. All 86 native backend tests, affected-package Clippy, 99 compiler library tests, 582 core tests, and 17 gates pass.

The controlled actual-worker profile rolling-wheel/typed-small-tensor-profile-2 measures Sim at 1.134402 seconds versus 1.347903 before this change (15.8% lower), with 1.12 seconds of user CPU. DAE and canonical Solve artifacts are byte-identical to the prior profile, as is the trace (SHA-256 f46cbcf4bf00620007139eb17416c2ece8c3ddf5665e577a3c4e22ef2dadc47c). Worker SHA-256 is 2f61244a39e14ed4b7c1f0c51d2b0a68e080dc28f4d6285e408fae7e39675c10. The first controlled profile reached its 180-second outer limit during source loading and measured no simulation; those artifacts remain retained.

The normal target/msl/multibody-typed-small-tensor-origin run measures Sim at 1.142239 seconds. Its first comparator step lacked OMC on PATH; using the pinned nix develop .#modelica shell compares the existing trace without rerunning Rumoca. One model is compared, all 184 trajectory and initialization channels high, with zero skipped/missing/nonidentifiable/deviating results. The fixed multibody-typed-small-tensor-canary preserves every phase, status, and band from multibody-manifold-selection-canary: nine compared models high, 175 initialization channels high, eleven unchanged refusals, and zero skipped/missing/nonidentifiable/deviating results. Receipt: rolling-wheel/typed-small-tensor-canary-delta.json; worktree digest: ad3a2ce8142fe8847c6f41d663005a5b9601ea229c7727364b87f8a00fee0608. The OMC speed target and combined quick/full remain outstanding. The latest complete cohort below predates these two native execution optimizations.

Previous focused work: select manifold derivative programs

RollingWheel’s three independent manifold blocks previously replayed all six directional programs for every state direction and discarded unrelated rows. Solve construction now binds each checked sparsity color to its exact program outputs, including shared tensor outputs. Execution calls each selected program once per color. Full residual certification and atomic state correction remain unchanged; native failures propagate without retrying the complete block. The governing owners are SPEC_0007’s Solve contract, SPEC_0032’s tensor ownership, and SPEC_0036/0039’s constructed output and sparsity proofs.

The reduced independent-block test fails before the repair: it observes four program executions where the two constraints require two (manifold-selection-red-2.log). Thirteen focused regressions pass afterward, covering permuted output placements, shared tensor programs, changing parameters, native/reference execution, and rollback after a later block fails. All 982 IR/evaluator/solver library tests, 99 compiler library tests, 582 core tests, 17 gates, formatting, and affected-package Clippy pass.

The controlled actual-worker Sim profile falls from 1.498229 to 1.347903 seconds (10.0%); the trace remains byte-identical with SHA-256 f46cbcf4bf00620007139eb17416c2ece8c3ddf5665e577a3c4e22ef2dadc47c. The normal target/msl/multibody-manifold-selection-origin gate measures Sim at 1.356090 seconds and compares one model: all 184 trajectory and initialization channels are high, with zero skipped, missing, nonidentifiable, or deviating results. This is still slower than OMC; differing timing boundaries preclude equating these measurements to OMC’s reported integration-only time.

The fixed target/msl/multibody-manifold-selection-canary retains every phase, status, and agreement band from multibody-causal-ready-canary: nine models compared high, all 175 initialization channels high, eleven unchanged refusals, and zero skipped/missing/nonidentifiable/deviating results. The receipt is rolling-wheel/manifold-selection-canary-delta.json, worktree digest a38f20cd00a5bed30dc189c4850eac38882b424ba6eab04ebbfd064d39610004. Profile artifacts are in rolling-wheel/manifold-selection-profile-1; worker SHA-256 is e2f9cd7b5dd61591b47bd0d8709fe0f1561b298e486d8193abb825b2d6454b60. This completes focused/canary validation. The latest complete cohort below predates this optimization; combined quick/full and the OMC speed target remain outstanding.

Previous focused work: bound construction cost

The RollingWheelSetDriving Solve timeout below was traced to repeated operation-prefix scans and repeated searches through causal candidates. A compact, source-bound destination-range index now supplies assignment producer lookups; a dependency queue preserves the exact earliest-ready causal order. Neither transformation expands tensor coordinates or changes numerical order. The normal target/msl/multibody-causal-ready-origin gate compares one model: all 892 trajectory and 892 initialization channels are high, with zero skipped, missing, nonidentifiable, or deviating results. Solve takes 8.149 seconds under the unchanged 10-second budget; Sim takes 1.206 seconds. This restores the originating focused case; the complete cohort confirmation is recorded below.

The fixed target/msl/multibody-causal-ready-canary has no phase, status, or band changes against multibody-pure-input-reuse-canary: nine models compared high, all 175 initialization channels high, zero skipped/missing/nonidentifiable or deviating results, and eleven unchanged refusals. The receipt is rolling-wheel/causal-ready-canary-delta.json, worktree digest 8681bf2a90c29ab9a8f0a860f95a5f9b783baef09a95fed4fc82a76ce6388ca2. The 519 IR/Solve/native and 162 structural library tests pass. Formatting, all-target/all-feature Clippy for the five affected compiler/runtime crates, and rumoca library/core/gate suites pass. The first Clippy attempt rejected a nested test closure; extracting its linear reference lookup fixed that lint.

The new controlled RollingWheel Sim profile measures 1.498 seconds and retains the identical trace. It collects 1,490 CPU samples with none lost. Native manifold Jacobian rows account for 14.37% of samples, algebraic projection rows 12.44%, and the full implicit Jacobian 1.75%; typed-call helpers add further work. These are self costs, not inclusive estimates. The OMC speed target remains unmet.

Latest focused work: reuse geometry without changing the trace

On top of 0dc3858bd6f0a8864f354df4560157395634c192, scalar Solve construction and AD share repeated unary arithmetic within each exact context. Native pure calls retain a result only under a construction-issued complete input coordinate; every input bit, including directional seeds, must match. Failure and assertion behavior remain covered by regressions.

RollingWheel’s focused Sim falls from 2.164 to 1.568 seconds. The separate actual-worker profile falls from 2.225 to 1.506 seconds, with an identical trace. Trig calls fall from 32,623,972 to 5,187,414. The normal focused gate target/msl/multibody-pure-input-reuse-origin-2 compares one model with all 184 trajectory and 184 initialization channels high, and zero skipped, missing, nonidentifiable, or deviating results. The first attempt hit a source-loading startup limit and measured no parity; its artifacts are kept.

The fixed target/msl/multibody-pure-input-reuse-canary has no phase, status, or band deltas against target/msl/multibody-auxiliary-primal-canary: nine models compared high, 175 initialization channels high, no skipped/missing/ nonidentifiable/deviating results, and eleven unchanged refusals. Worktree digest: e333300748c075ba3708f0749572e75e452536c98726a4e6d610e4603c73b8dc. The exact receipt is rolling-wheel/pure-input-reuse-canary-delta.json.

The OMC performance gap remains open. Details and reduced RED/GREEN evidence are in the generated-kernel comparison.

Latest complete measurement

target/msl/multibody-causal-ready-full passes at commit 4585575a610fd989625d2cd12fa63df398c70247. The complete 566-model comparison measures 157/566 strict-high (27.74%): 157 models compared, eighteen reviewed exclusions skipped, zero missing/nonidentifiable traces, and zero deviating channels. All 19,412 initialization channels are high. MultiBody is 21/42 high. Every previously high model stays high; RollingWheelSetDriving regains high parity under the unchanged Solve budget. The sole band change is that restoration. Ten phase/status rows change: six unpassed Fluid cases now report harness failures, and three existing simulation refusals reach typed structural failures instead of timing out. Those changes earn no parity credit.

The gate takes 463.64 seconds with eleven Rumoca simulation workers; the OMC reference/comparison stage takes 313.85 seconds. Host I/O pressure was substantial during trace handling, so this gate duration is not a solver benchmark. The per-model receipt is rolling-wheel/causal-ready-full-delta.json; comparator SHA-256 is ba7faf4f96e4f2c38e1a8c76ef10ddb4ccc0cda74e35f2652690c50b7e098e53. Worktree digest is 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35; the preserved untracked communication file accounts for the dirty marker. Combined quick/full and the OMC speed target remain outstanding.

Previous complete measurement: exact pure-call input reuse

target/msl/multibody-pure-input-reuse-full passes the checked-in quality gate at commit 716bba8bef01f04259655e88dcb0e6112b66011a in 185.80 seconds. The comparator measures 156/566 strict-high (27.56%): 156 models compared, eighteen reviewed exclusions skipped, zero missing/nonidentifiable traces, and zero deviating channels. All 18,520 initial-condition channels are high. MultiBody remains 20/42 high, with a changed membership.

Against multibody-seed-linearization-full, Rectifier, CompareTransformers, and DoublePendulumInitTip advance from refusal to strict-high. RollingWheelSetDriving loses its previous strict-high result: the worker exceeds the 10-second Solve construction budget before simulation. The aggregate gain does not excuse that regression. The newer complete run above confirms its repair; no baseline promotion is claimed. Fourteen models change phase/status details in total. The exact receipt is rolling-wheel/pure-input-reuse-full-delta.json.

The worktree digest is 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35; the comparator artifact SHA-256 is dbc9374983f1918a29fb78850d05593a7a3becee194c72d0c000789727fe7b56. The tracked source was committed; the preserved untracked communication file accounts for the dirty-worktree marker. Combined quick/full remain pending.

Previous complete measurement: shared seed linearization

target/msl/multibody-seed-linearization-full passes at commit 5fbcc036d50f1124b99069797a64a41fcd500d00, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The full 566-model gate takes 153.51 seconds with eleven Rumoca workers. All 154 compared models remain strict-high (27.21%), with eighteen unchanged reviewed exclusions skipped, zero missing/nonidentifiable traces, and zero deviating channels. All 18,693 initial channels remain high. MultiBody stays 20/42 high. Every model retains its previous phase outcome and agreement band; the exact delta is rolling-wheel/seed-linearization-full-delta.json.

The isolated profile measures RollingWheel Sim at 1.248 seconds, versus 1.309 before fixed-coordinate matrix reuse, with a byte-identical trace. The critical OMC performance gap remains open. Combined verify quick and verify full remain pending; no PR, release, or baseline promotion is claimed.

The next structural investigation compares the requested and selected state coordinates. Flat and the checked DAE retain StateSelect.always on wheel1.x, wheel1.y, wheel1.angles, and wheel1.der_angles, but classify these eight coordinates as algebraic. Rumoca integrates twelve other coordinates; OMC integrates those eight requested coordinates. This establishes a selection difference, not yet its causal contribution to the runtime gap. The 21-variable dynamics block versus OMC’s six-variable block remains a related lead, recorded in rolling-wheel/algebraic-kernel-next-target.json.

Previous complete measurement: on-demand derivatives

target/msl/multibody-on-demand-derivatives-full passes at commit c1145fa9a50e0af44e4bc4ac02c74a5317bd14f9, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The full 566-model gate takes 137.07 seconds with eleven Rumoca workers. All 154 compared models remain strict-high (27.21%), with eighteen unchanged reviewed exclusions skipped, zero missing/nonidentifiable traces, and zero deviating channels. All 18,693 initial channels remain high. MultiBody stays 20/42 high. Every model retains its previous phase outcome and agreement band; the exact delta is rolling-wheel/on-demand-derivatives-full-delta.json.

The isolated profile below improves RollingWheel Sim from 1.640 to 1.309 seconds with a byte-identical trace. The critical OMC performance gap remains open. Combined verify quick and verify full remain pending; no PR, release, or baseline promotion is claimed. Next: repeated algebraic matrix work at fixed directional-derivative coordinates and source-backed elimination in the dynamics block. rolling-wheel/algebraic-kernel-next-target.json records the comparison with OMC’s six-variable system: Rumoca retains 21 unknowns, and its existing tearing only reduces that to fourteen. That is a structural lead, not proof that changing the solve dispatch preserves the earlier electrical fixes.

Previous complete measurement: deferred observation evaluation

target/msl/multibody-deferred-observation-full passes at commit 9639fad96b733e94fb4aadc454cdb04e0eedcf55, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The full 566-model gate takes 152.48 seconds with eleven Rumoca workers. All 154 compared models remain strict-high (27.21%), with eighteen existing reviewed exclusions skipped, zero missing/nonidentifiable traces, and zero deviating channels. All 18,693 initial channels remain high. MultiBody stays 20/42 high; every preceding high model is retained.

RollingWheel takes 1.657 seconds in concurrent Sim and RollingWheelSetDriving takes 2.060 seconds. The isolated RollingWheel profile below establishes the performance improvement separately. ThyristorBridge2Pulse_RLV_Characteristic now completes in 11.181 seconds under the unchanged budget; its pre-existing exclusion concerns insufficiently converged OMC period-aggregate reference values, so it earns no high-parity credit. This accounts for the eighteenth skipped completion. Inverse_sh_TX returns its previous Flatten refusal; no model changes agreement band. Exact deltas are in rolling-wheel/deferred-event-left-full-delta.json.

All seventeen repository inspection gates also pass, including spec budgets, source citations, and production file-size checks (deferred-event-left-repo-gates-1.log).

The critical RollingWheel runtime gap remains open. Combined verify quick and verify full remain pending; no PR, release, or baseline promotion is claimed. The next measured target is proactive completed-step derivative evaluation, followed by repeated algebraic linearization work.

Previous complete measurement: grouped Jacobian outputs

target/msl/multibody-grouped-jvp-full passes at commit 2c637488b3c88703d4f729eb30f7d6453b58543e (compiler change b51acb48), working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The full 566-model gate takes 147.38 seconds with eleven Rumoca workers. All 154 compared models remain strict-high (27.21%); seventeen unchanged reviewed exclusions remain skipped, with zero missing/nonidentifiable traces and zero deviating channels. All 18,693 initial channels remain high. MultiBody remains 20/42 high, and every previous high model is retained.

RollingWheel takes 2.209 seconds in concurrent Sim; RollingWheelSetDriving takes 2.803 seconds. The isolated RollingWheel profile below remains the performance evidence, and its critical OMC gap remains open. No model changes agreement band. Three absent models change failure classification: Inverse_sh_TX reaches DAE construction refusal, BevelGear1D reports a runtime contract failure instead of timeout, and GenerationOfFMUs returns its previous structural refusal. Exact deltas are in rolling-wheel/grouped-jvp-full-delta.json.

Combined verify quick and verify full remain pending. The immediate priority is RollingWheel execution cost; no coverage expansion, PR, release, or baseline promotion is claimed.

Previous complete measurement: affine seed removal

target/msl/multibody-affine-seed-full passes the local fallback gate at commit 9fb1b2476977ea27225287751e2d8adbb4afafea, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The full 566-model comparison takes 152.98 seconds with eleven requested, admitted, and pinned Rumoca workers. All 154 compared models remain strict-high (27.21%), with seventeen unchanged reviewed exclusions, zero missing/nonidentifiable traces, and zero deviating channels. All 18,693 initial channels remain high. Every prior high model, including all nine electrical counterexamples and DCPM_Cooling, is retained; MultiBody remains 20/42 high.

RollingWheel retains 184/184 high channels and the same maximum channel bounded normalized L1 1.0279228736061218e-4. Concurrent Sim takes 2.394 seconds, with 3.887 seconds of separate source/Solve/backend preparation. RollingWheelSetDriving retains 892/892 high channels and takes 4.385 seconds in Sim. These cohort timings are not isolated benchmarks; the clean profile pair below measures the seed-removal change separately.

No model changes agreement band. GenerationOfFMUs remains absent: this run hits the unchanged ten-second Solve budget before returning the previous structural EL005 refusal. That phase outcome is retained in rolling-wheel/affine-seed-full-delta.json; it earns no coverage credit. The complete MSL gate is green, but the critical RollingWheel performance gap remains open. Combined verify quick and verify full remain pending while that compiler performance work continues. No PR, release, or baseline promotion is claimed.

Previous complete measurement: affine tensor functions

target/msl/multibody-affinity-full passes the local fallback gate at commit aecb01dc7038c7cedede5fe6dbc85c62b0c3c74e, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The complete 566-model run takes 139.77 seconds with eleven requested, admitted, and pinned workers. All 154 compared models remain strict-high (27.21%), with seventeen unchanged reviewed exclusions, zero missing or nonidentifiable traces, and all 18,693 initial channels high. There are zero deviating trajectory channels. Every prior high model, including the nine electrical counterexamples and DCPM_Cooling, remains high. MultiBody stays 20/42 high.

RollingWheel’s changed trace passes with 184/184 channels high, maximum channel bounded normalized L1 1.0279228736061218e-4. Its concurrent Sim time is 3.597 seconds; source/Solve/backend preparation is separately 3.722 seconds. RollingWheelSetDriving retains 892/892 high channels and takes 4.850 seconds in Sim. No model changes band. The sole phase change is GenerationOfFMUs, which remains absent and returns its earlier structural EL005 refusal rather than the preceding run’s Solve-stage timeout. The complete delta is rolling-wheel/affinity-full-delta.json.

The isolated profile pair below supports the RollingWheel improvement; this cohort timing is not a head-to-head benchmark. The critical runtime gap remains open. Combined verify quick and verify full have not run for this change; no PR, release, or baseline promotion is claimed.

Previous complete measurement: prepared algebraic refresh

target/msl/multibody-prepared-stages-full passes the local fallback gate at commit 440f4466fec6987eb1cbcb6c68f34500566319ae, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The complete 566-model run takes 196.38 seconds with eleven requested and admitted workers; the OMC reference/comparison stage accounts for 91.87 seconds.

All 154 compared models are strict-high (27.21% of 566), with seventeen unchanged reviewed exclusions, zero missing/nonidentifiable traces, and all 18,693 initial channels high. There are zero deviating trajectory channels. Every previous high model remains high, including the nine electrical counterexamples. MultiBody remains 20/42 high.

DCPM_Cooling advances from a simulation timeout to 7.911 seconds in Sim, with all 195 channels high and maximum channel bounded normalized L1 1.518715114241044e-5. This is the only band change. The other phase change is GenerationOfFMUs, which remains absent: its previous structural refusal is replaced by a ten-second Solve-stage timeout in this run. The complete delta is rolling-wheel/prepared-stages-full-delta.json.

RollingWheel takes 4.556 seconds in Sim and RollingWheelSetDriving 5.015 seconds; all 184 and 892 respective channels remain high. The clean isolated RollingWheel pair below remains effectively unchanged at 4.283 versus 4.303 seconds. The critical OMC runtime gap remains open. Combined verify quick and verify full have not run for this change; no PR, release, or baseline promotion is claimed.

Previous complete measurement: selected native residual execution

target/msl/multibody-native-residual-full passes the local fallback gate at commit a2cd4deae1c50e9650c0ca3124452c0d72013b2b, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The complete 566-model run takes 120.83 seconds with eleven requested and admitted workers. OMC reuses all 170 existing reference results.

All 153 compared models remain strict-high (27.03% of 566), with seventeen unchanged reviewed exclusions, zero missing/nonidentifiable traces, and all 18,498 initial channels high. There are zero deviating trajectory channels. All 566 phase/simulation outcomes and bands match the native-manifold full run; all nine electrical counterexamples retain their high bands. MultiBody remains 20/42 high. rolling-wheel/native-residual-full-delta.json records the complete comparison and artifact hashes.

RollingWheel takes 4.380 seconds in Sim and RollingWheelSetDriving 4.927 seconds; build times are separately 4.188 and 10.873 seconds. The isolated profile pair below establishes the selected-residual dispatch improvement. The OMC gap remains a critical runtime bug. The fixed 145-model median speed ratio is 2.635 and the expanded 170-model median is 1.730; this concurrent run with cached OMC timings does not establish an isolated cohort speedup. Combined verify quick last passed before ten subsequent implementation changes; verify full has not run. No PR, release, or baseline promotion is claimed.

Previous complete measurement: native manifold execution

target/msl/multibody-native-manifold-full passes the local fallback gate at commit 55acfbbdbef06c7d3b86820ce3e2b71deaa5cf40, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The complete 566-model run takes 133.14 seconds with eleven requested and admitted workers. OMC reuses all 170 existing reference results.

All 153 compared models remain strict-high (27.03% of 566), with the same seventeen reviewed exclusions, zero missing/nonidentifiable traces, and all 18,498 initial channels high. There are zero deviating trajectory channels. All 566 phase/simulation outcomes and bands match the preceding full run; all nine electrical counterexamples retain their high bands. MultiBody remains 20/42 high. rolling-wheel/native-manifold-full-delta.json records the complete comparison and artifact hashes.

RollingWheel takes 5.215 seconds in Sim and RollingWheelSetDriving 5.423 seconds; build times are separately 4.375 and 10.568 seconds. The isolated profiles below establish the native-manifold dispatch improvement. The remaining OMC gap is still a critical runtime bug. The fixed 145-model median speed ratio is 2.164 and the expanded 170-model median is 1.463; this concurrent run with cached OMC timings does not establish an isolated cohort speedup. Combined verify quick last passed before nine subsequent implementation changes; verify full has not run. No PR, release, or baseline promotion is claimed.

Previous complete measurement: state-only manifold projection

target/msl/multibody-state-only-manifold-full passes the local fallback gate at commit f485ef13f2e3a79c9ae136c193eb3560f05114dc, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The complete 566-model sweep takes 175.94 seconds. Eleven Rumoca workers were requested; available memory limited execution to ten. OMC reuses 169 cached results and runs one new model in one persistent session. The run metadata’s 14-worker OMC setting describes its configured budget, not fourteen active new simulations.

All 153 compared models are strict-high (27.03% of 566), with seventeen reviewed exclusions, zero missing/nonidentifiable traces, and all 18,498 initial channels high. No compared trajectory channel deviates. All nine electrical counterexamples retain their high bands. MultiBody returns to 20/42 high; RollingWheelSetDriving is the only phase/simulation and band change from multibody-singleton-degree-full. Every preceding high model remains high. The delta, electrical checks, and artifact hashes are in rolling-wheel/state-only-manifold-full-delta.json under the campaign directory.

The wheel-set build still takes 10.708 seconds, separately from simulation. Its complete trace is byte-identical to the earlier singleton-degree diagnostic (3dc2333f4b9d6d69bead292633f3304aa57204f430040d50c00fb62753006a5b), with maximum channel bounded normalized L1 5.061492181479045e-6. RollingWheel’s trace changes slightly and remains all-high, maximum channel bounded normalized L1 1.1053709819053231e-4. OMC’s cached RollingWheel simulation timing is 0.134 seconds: substantial runtime overhead remains. The fixed 145-model performance cohort median is 2.585 versus 2.245 before; the expanded 170-completion median is 1.688. Memory pressure, concurrency, and cached OMC timings prevent interpreting this pair as an isolated speedup.

The runtime repair and its focused/canary checks are complete. The remaining 22 MultiBody models still need investigation. Combined verify quick last passed before eight subsequent implementation changes; verify full has not run. No PR, release, or baseline promotion is claimed.

In progress: requested states and repeated tensor reduction

The candidate based on eb325ac7173ddd0757783f6c3ed5cc77e894c251 promotes continuous Real declarations marked StateSelect.always through one checked whole-DAE reconstruction before matching. MLS §4.8.7.1 and STRUCT-T07 govern this selection; SPEC_0032 governs preservation of tensor owners. The initial reduced alias model selects q[1:2] before the change and x[1:2] after it, agreeing with OMC. Output causality, parameter roles, assertions, and contradictory fixed initial values have focused regressions.

Promotion alone regressed RollingWheel to structural refusal (rolling-wheel/requested-states-origin-1). The actual reduction recorder shows the demoted path reaching one unmatched der(wheel1.y). A previous component replacement had wrapped the position residual in ArrayUpdate; the next component proof did not read through that checked update to its untouched y component. The pristine retry was less reduced and is not evidence that the earlier angular-state demotion failed.

The six-equation SelectedComponents regression reproduces the same gap with der(p)=-p; p={x,y,z} and requested x/y states. It fails structural preparation before the component-selection repair (requested-components-red-1.log). Following the exact literal, full-rank update selection makes RollingWheel structurally regular (requested-states-reduction-probe-2.log). The prepared DAE retains all eight requested coordinates plus body velocity and joint position tensors: fourteen integrated coordinates versus OMC’s eight. Complete elimination of these dependent coordinates remains open.

The reduced model then exposed Solve derivative extraction failing on the updated tensor (requested-components-green-1.log). Reusing its existing constant update selector repairs that boundary. Both integrators now reproduce all six analytic channels (requested-components-green-2.log). OMC also simulates this exact source, selecting x, y, and p[3]; its maximum analytic channel error is below 1e-5 (requested-components-omc-1).

The next actual-model gate, target/msl/multibody-requested-components-origin, still fails Solve and compares zero models. Differentiating the first no-slip constraint produces a sum of products equal to zero, but the derivative extractor required a subtractive root. A second reduced model, der(q)=-q; 0=x*q+1, with x requested, reproduces that error (requested-product-red-1.log). Passing the complete residual to the existing scalar affine proof makes both integrators agree with its analytic solution (requested-product-green-1.log). OMC also simulates the exact product fixture successfully (requested-product-omc-1/comparison.json).

RollingWheel remains refused after this repair (requested-components-lowering-probe-3.log). The equivalent dot-product fixture 0={x,1}*{q,1} also reproduces this refusal (requested-dot-product-red-1.log). Inspecting the actual prepared BLT shows a more fundamental problem: block 484 contains three angular accelerations and 21 algebraics (requested-state-blocks-probe-1.log). Solve drops the derivative entries from the algebraic projection and then tries to isolate those derivatives separately. A scalar pivot need not be invertible even when this joint system is regular.

The proposed STRUCT-T09 normalization introduces exact whole-tensor algebraic aliases for implicit derivatives, substitutes their typed coordinate reads, and appends der(state)=alias. The joint equations then remain one compiled algebraic system. MLS Appendix B and §8.3.1 justify the equivalent extension; §8.6 initialization obligations and existing runtime rank/convergence checks remain required. No coefficient is estimated by subtracting residual samples. The normalization is implemented and its originating-model gate now completes: target/msl/multibody-requested-aliases-origin compares one model, with all 184 channels and initialization channels high and zero skipped, missing, nonidentifiable, or deviating traces. This focused result is not a new cohort measurement. Its isolated Sim time is 2.324479 seconds, slower than the prior 1.248-second candidate. The generated-C comparison and diagnostic counters identify substantial repeated geometry evaluation; see RollingWheel kernel comparison.

The new mixed derivative/dot-product reproduction fails before normalization (requested-mixed-dot-red-1.log) and passes both integrators afterward (requested-aliases-green-2.log). The same source also passes OMC, with maximum analytic invariant error below 4e-6 (requested-mixed-dot-omc-1/comparison.json). All six requested-state regressions pass (requested-aliases-selected-tests-2.log). Broader validation exposed native tensor linear-solve, explicit derivative family, and scalar affine paths being unnecessarily replaced by aliases. The selection now retains those owners, using exact matching and compact family bodies. Outside mixed BLT blocks, admission requires a derivative consumed through a tensor expression. Scalar-only equations keep their existing affine construction and rejection checks. No existing regression was weakened.

All 961 library/core tests pass (requested-aliases-libraries-4.log), as does all-target/all-feature Clippy for structural, Solve, and the public compiler crate (requested-aliases-clippy-1.log). Temporary compiler probes were removed. The fixed canary target/msl/multibody-requested-aliases-canary has no phase, simulation-status, or comparison-band changes against target/msl/multibody-requested-states-canary: nine compared models are high, all 175 compared initialization channels are high, and skipped, missing, nonidentifiable, and deviating counts are zero. The eleven remaining targets retain their prior refusals. The candidate worktree digest is 3dd6c3edbc76ce9f91b6cfb4c4f3203d91c8e29e72c84e93d42cd187b231dc39; requested-aliases-canary-delta.json records the exact comparison and artifact digests. This is Tier 1 regression evidence only. The final originating-model rerun after native-owner preservation, target/msl/multibody-requested-aliases-final-origin, also compares one model with all 184 channels and 184 initialization channels high, zero skipped, missing, nonidentifiable, or deviating traces, and Sim time 2.358592 seconds.

Combined quick/full and a new full cohort gate remain pending while the RollingWheel kernel work continues. The original-model refusal is closed by the focused comparison; the performance regression remains open. Private logs in this section are under .git/multibody-campaign/rolling-wheel/.

Inspect the actual structurally transformed equations

The worker wrote the pre-transform input into ir-structural-dae.json and ir-structural-dae.mo. This made an apparent DAE-to-DAE comparison compare the same input twice. The build observer now borrows the immutable LoweredSolveModel pair already retained by phase-solve and serializes its prepared_dae(). No structural pass is repeated, and no numerical program or simulation calculation changes. A failed lowering leaves no structural artifact; recognized prior structural files are cleared before a new request.

The focused source model der(x)=v; der(v)=a; v=-x, with x initially fixed, reduces from two states to one. Before the fix its structural artifact still names both (structural-artifact-red-2.log); afterward the artifact and Solve both retain only x. The first RED attempt was a fixture compilation error. A second regression checks that failed reduction leaves neither prior nor mislabelled structural products while preserving unrelated files. Its initial fixture failed too early in ToDae; the corrected balanced singular fixture reaches the intended structural failure (structural-artifact-green-2.log).

All 186 simulation, worker, parameter-override, and legacy simulation-worker tests pass (structural-artifact-libraries-1.log). The obsolete import noted by that run was removed; all-target/all-feature Clippy for the four affected crates and rustfmt pass (structural-artifact-clippy-1.log). The fixed multibody-structural-artifact-canary retains all twenty phase and band outcomes against multibody-seed-linearization-canary: nine compared models and 175 initial channels high, zero skipped/missing/nonidentifiable/deviating traces. Source: 40a22015 plus digest 905e884225824a8e240af7db6c6b1d69b028a79547ac099698336b15c893642a; receipt: rolling-wheel/structural-artifact-canary-delta.json.

structural-artifact-origin-1 records the actual RollingWheel transformation: 3,199 input expressions become 4,024, with all 255 continuous equation owners retained. The body-frame position changes from state to algebraic; the twelve remaining state coordinates and all eight unselected always coordinates are now verified in the transformed DAE. Canonical Solve JSON and the complete trace remain identical to the preceding profile. Trace SHA-256: 61e23f642f39489afa0717c767501558972f2e29419b77b088befbe8b2aff3dd. The diagnostic profile takes 1.265 seconds in Sim with 237 samples; this is no performance improvement. Worker SHA-256: dc0a892657723018dfdc4685e32a549ca7cf1dd97debe5c94ff3c1fc29f28019.

The reduced PreferredTensorState model has x=q; der(q)=-q, with whole array x[2] marked StateSelect.always and q[2] marked avoid. Rumoca keeps q as its state array through structural lowering; OMC selects x[1:2]. This rules out lost parameter binding or a MultiBody-specific cause. The MLS StateSelect contract and SPEC_0007/STRUCT-T07 place the missing selection in structural lowering. No state-promotion fix or causal runtime benefit is claimed yet. Source, actual stage dumps, OMC generated equations, and the next proof obligation are retained in rolling-wheel/state-selection-diagnosis.json. These small diagnostic runs are not cohort parity or comparative performance measurements; their process executions briefly overlapped, so their timings are not used for comparison.

Retain algebraic linearizations across directional seeds

The state-only Jacobian path already retains its settled algebraic coordinate, but project_algebraic_seed_with_plan_inner rebuilt and factorized every selected block for every direction. Its Y/P/time point is unchanged across successive seeds. The focused nonlinear tensor-output regression records a second matrix evaluation for the second seed (seed-linearization-red-3.log). The preceding RED attempts exposed an incomplete fixture and a counter that missed the scalar reverse path; neither is the behavioral reproduction.

Derived structural construction now certifies repeatability from the complete primal and both directional output catalogs, including otherwise discarded effects. The runtime retains the same block matrix, row scales, and LU for that exact complete Y/P/time coordinate. Constructor-issued block indexes keep the matrices attached to their original systems. Each requested seed still propagates its own right-hand side and passes the original residual check. Coordinate changes invalidate retained matrices, failure clears the entries and restores the caller’s seed, and cloned runtimes start with empty numerical caches. The model’s table/call context is immutable. Neither canonical programs nor tensor structure changes; SPEC_0038 records the numerical reuse condition.

Regressions cover state and algebraic values, parameter and time changes, adjacent state bits, signed-zero time, parameter seeds, a singular coordinate, missing construction proofs, clone isolation, and impure or ambiguous primal/JVP output owners. All 1,110 tests across Solve IR, evaluation, solver, simulation, BDF, and RK45 pass (seed-linearization-libraries-1.log), together with their all-target/all-feature Clippy (seed-linearization-clippy-1.log). All seventeen repository inspection gates pass (seed-linearization-repo-gates-2.log); the first pass caught SPEC_0038 twelve words over budget, resolved by shortening the new rule without changing it.

multibody-seed-linearization-canary preserves all twenty phase and band outcomes against multibody-on-demand-derivatives-canary. Its nine compared models and 175 initialization channels remain high, with no skipped, missing, nonidentifiable, or deviating traces. Source: 24c1d944 plus digest 7f0bb9aae11fcef6efc474ae27774035517cde8c1be898636d9a512d69ffe1d4; the exact delta is rolling-wheel/seed-linearization-canary-delta.json.

The isolated critical-seed-linearization-profile-1 measures RollingWheel Sim at 1.248 seconds against 1.309 previously; user CPU decreases from 1.30 to 1.23 seconds. The roughly 4.7% change is a single-pair measurement. The complete trace is byte-identical, SHA-256 61e23f642f39489afa0717c767501558972f2e29419b77b088befbe8b2aff3dd, and canonical Solve JSON is unchanged. Worker SHA-256: 7b4d386c05122e0f57dbacb1085b4ff495d361ec2fe80e28d00267820bbd49db. The profile retains 236 CPU samples with zero lost samples. Residual/JVP program 223 and trigonometric calls remain prominent. This modest change does not close the critical OMC gap. The complete multibody-seed-linearization-full gate passes at 5fbcc036 with all 566 phase and band outcomes unchanged; see the latest complete measurement above.

Request derivatives only when the numerical method needs them

SolveMeKernel::completed_integrator_step evaluated state derivatives solely to warm its private accepted-point cache. This executes the derivative algebraic closure even when the numerical method has no derivative request at that point. The component now performs only its required event/history work and cache invalidation during completion. The ordinary derivative getter still evaluates and caches requested points. SPEC_0038 records this ownership rule; no equation, integration tolerance, or required observation changes.

The focused harmonic-oscillator regression uses a counting native-interface adapter executing the real prepared residual. Before the fix the first completion makes one unrequested RHS call (on-demand-derivatives-red-2.log). Afterward completion makes none, explicit requests return the expected derivatives, and identical requests reuse their cache across successive steps. The first attempted RED exposed a fixture type error and is not behavioral evidence. All 607 solver, simulation, BDF, and RK45 tests pass, as does all-target/all-feature Clippy (on-demand-derivatives-libraries-2.log, on-demand-derivatives-clippy-1.log). All seventeen repository inspection gates pass (on-demand-derivatives-repo-gates-1.log). The first library build exposed the now test-only allocating derivative wrapper; production uses the existing borrowed slice operation.

The fixed multibody-on-demand-derivatives-canary preserves all twenty phase and agreement-band outcomes against multibody-deferred-observation-canary. Its nine compared models and 175 initialization channels remain high, with no skipped, missing, nonidentifiable, or deviating trace. Source: e6017e33 plus working-tree digest ff2f7b90df633432c6fa82da8d08a77d5558d48c8f3c8169d4cac837f374c542; the exact delta is rolling-wheel/on-demand-derivatives-canary-delta.json.

The isolated critical-on-demand-derivatives-profile-1 improves RollingWheel Sim from 1.640 to 1.309 seconds, with user CPU decreasing from 1.61 to 1.30 seconds. This single pair supports an approximately 20% improvement. The trace remains byte-identical, SHA-256 61e23f642f39489afa0717c767501558972f2e29419b77b088befbe8b2aff3dd, and canonical Solve JSON is unchanged. Worker SHA-256: 23b491debb8eb362a357762d42d5d3a46ac3f94acc1dae24f590caa8c070405a. The profile retains 255 CPU samples with zero lost samples. Algebraic JVP and typed pure-call directional programs remain prominent; this is still much slower than the existing OMC reference. The subsequent complete cohort passes at c1145fa9, as recorded above; no coverage increase or release readiness is claimed.

Defer speculative event-left outputs

An actual-worker census at 94d0d2ce identifies a host-level source of unnecessary algebraic work. RollingWheel takes 1,432 accepted steps, 3,571 derivative-kernel calls, and 444 directional calls. The full algebraic refresh runs 1,932 times: 500 requested positive-time samples plus an unconditional event-left output candidate for every accepted step. Uneventful endpoints discard those candidates. OMC’s repeated reference runs take 1,028 steps and 1,329 ODE calls; step count alone cannot explain the runtime gap.

commit_accepted_endpoint now retains the admissible left coordinate, native sampled states, and opaque pre-callback FMU snapshot. It evaluates outputs only when a known time event or returned step event needs publication. Evaluation restores that snapshot, uses the ordinary output operation, and restores the current complete component state on success, failure, or unwind. No derivative, root, event, or required observation is omitted. SPEC_0038 and SPEC_0044 §6 record this host rule; FMI 3.0.2’s state restore and completed-step semantics remain the component boundary.

The reduced RED has a valid endpoint and an unrequested interior observation whose algebraic branch is undefined. The old host fails solely on that discarded candidate (deferred-event-left-red-1.log); the repaired host advances. A known time event still propagates the same getter failure. Separate regressions prove saved-input evaluation, complete current-state restoration after errors and panic, original panic propagation, and foreign-snapshot rejection. All 606 solver, BDF, RK45, and simulation tests pass, as does all-target/all-feature Clippy (deferred-event-left-libraries-2.log, deferred-event-left-clippy-1.log).

The fixed multibody-deferred-observation-canary takes 47.10 seconds, preserving all twenty phase/band outcomes. All nine compared models and 175 initialization channels remain high; no trace is skipped, missing, nonidentifiable, or deviating. The measured source is 94d0d2ce plus digest b18a0e7f309402e31be3b624c5510115e4eefb7f8d261f8e9b63daaf774eff1e.

The clean critical-deferred-observation-profile-1 run improves RollingWheel Sim from 2.138 to 1.640 seconds (user CPU 2.11 to 1.61 seconds). The complete trace remains byte-identical, SHA-256 61e23f642f39489afa0717c767501558972f2e29419b77b088befbe8b2aff3dd, and canonical Solve JSON is unchanged. The native worker SHA-256 is 22a5ba80fa1bde722f5898d9f1b59d468592ba288721372345d1b2f228666d2a; 309 CPU samples have zero lost samples. This single isolated pair supports an approximately 23% improvement, not parity with OMC. Temporary probes are removed; kernel-census-probes preserves their patch, original files, and manifest. The subsequent full cohort passes at 9639fad9, as recorded above. The critical performance gap remains open.

Reuse all tensor JVP outputs within a prepared color

The selected-row interface recomputed an entire JVP program to return each requested output. The earlier actual RollingWheel census recorded 40,821 evaluations of each of program 195’s three selected outputs. A three-column Jacobian therefore invoked the same tensor program nine times. OMC’s generated linear-system matrix functions fill their selected entries directly; they do not replay a whole tensor derivative separately to retrieve each component.

Under SPEC_0007 SOLVE-C56, derived structural-artifact construction now binds each projection color to ordered program/output groups in both solver-Y and solver-Y/parameter seed spaces. It uses the checked logical output catalogs, retains their nonidentity placement, and does not rewrite the scalar/tensor programs. The runtime consumes those groups at one immutable numerical point, omits reverse-completed rows, and retains all selected outputs of each native or interpreted program invocation. Native failures propagate without a retry. There is no persistent numerical cache or runtime grouping by hashes, pointers, or operation equality.

The construction also checks operation repeatability exhaustively, including compact nested branches and folds. Impure random operations prevent reuse even when their values are discarded. Ambiguous or missing logical output ownership cannot issue a grouped selection. Table access retains the same immutable external-table context for each invocation.

rolling-wheel/grouped-jvp-red-1.log records the coupled-system regression: the numerical solution was correct, but the native interface received four single-output calls instead of two grouped calls. The repaired regression reuses the runtime while changing the matrix coefficient through 2,-3,0,2. grouped-jvp-effects-red-1.log separately captures the initially missing guard for an unused impure operation. Regressions cover both seed-space maps, reverse-completed rows, input extents, unrelated-program isolation, table failures, and actual JIT call counts. All 307 Solve-IR, 190 evaluator, 78 native, 463 solver, and 128 simulation tests pass (grouped-jvp-libraries-2.log). All-target/all-feature Clippy for those crates passes (grouped-jvp-clippy-3.log); its earlier two attempts exposed nesting only.

The fixed target/msl/multibody-grouped-jvp-canary preserves all 20 phase and agreement-band outcomes against multibody-affine-seed-canary. Its nine compared models and 175 initialization channels remain high, with zero skipped, missing, nonidentifiable, or deviating traces. Eleven requested, admitted, and pinned workers ran the canary. The delta receipt rolling-wheel/grouped-jvp-canary-delta.json binds dirty-tree digest b8a0a484309d158bbaac8d2cbee7ba859808690a379cd7a918934abd4708e1b2 to parent 52d578f101f0f992eebc9e7b4bf15366a8acab54.

The clean isolated rolling-wheel/critical-grouped-jvp-profile-1 takes 2.138 seconds in Sim and 2.11 seconds of user CPU, versus 2.315 and 2.28 seconds in critical-affine-seed-profile-1. This single pair supports roughly 8% less runtime. The trace is byte-identical (61e23f642f39489afa0717c767501558972f2e29419b77b088befbe8b2aff3dd), and the entire emitted canonical Solve JSON compares equal. The profile retains 414 samples with zero loss and named JIT symbols. Program 223’s residual and JVP, plus their trigonometric work, remain material costs; the OMC gap remains open. rolling-wheel/grouped-jvp-diagnosis.json pins the evidence and worker hashes.

Omit singular guesses before prepared affine solves

RollingWheel’s positive-time runtime census found 5,540 failures of the optional scalar seed for wheel1.rollingWheel.delta_0[3] in projection block 823. Each failure restored the incoming snapshot and abandoned the staged refresh for the complete projection plan. The earlier census recorded 2,015,815 block evaluations, including 13,607 executions of the angle-rate block. These are instrumented diagnostic counts, not performance measurements.

The preserved tensor equations locate the first divergent producer precisely: the seed isolates delta_0[3] from dot(delta_0, e_long_0) = 0. The source defines e_long_0 = normalize(cross({0,0,1}, e_axis_0)), so that scalar pivot is zero. The coupled contact equations use the orthogonal directions e_axis_0, e_long_0, and their cross product; the complete matrix remains nonsingular within the source assertion domain. OMC’s generated linear system 664 also retains a coupled three-variable solve, using body height in place of delta_0[3]. The equations and seed owner are pinned in rolling-wheel/affine-seed-diagnosis.json.

Under SPEC_0007 SOLVE-C56, ContinuousRefreshOwners now omits projection seeds after deriving affinity from the canonical residuals and before emitting exact assignment schedules. The same construction applies to algebraic, derivative, root, event, clock, and remainder owners. An affine solve already computes its solution independently of the incoming guess. Nonlinear blocks retain their seeds, and the original tensor residuals, assertions, coupled matrices, and coordinate refinement remain authoritative.

The reduced regression uses y - k*x = 1, x + y = 3: scalar isolation of x divides by zero at k=0, while the coupled determinant is -(k+1). affine-seed-red-1.log records the failing schedule assertion. The repaired test reuses the prepared runtime across k=0,1,3,0; companion tests cover nonlinear seed retention and every owner/remainder’s emitted schedule. All 303 Solve-IR, 190 evaluator, 459 solver, and 128 simulation tests pass (affine-seed-libraries-{1,2}.log), as does all-target/all-feature Clippy for these crates (affine-seed-clippy-1.log). Temporary probes were removed.

The fixed target/msl/multibody-affine-seed-canary preserves all 20 phase and band outcomes against multibody-affinity-canary: all nine compared models and 175 initialization channels remain high, with zero skipped, missing, nonidentifiable, or deviating results. The run used 11 requested/admitted/pinned workers; rolling-wheel/affine-seed-canary-delta.json binds its dirty tree digest 4e0d8507938500c5264e1c551dbb49281b070abb631e9bf27c53fd0554f99056 to parent 2db3389d45583ba3551b1cb19124a75eeef2ef0c.

The clean isolated profile rolling-wheel/critical-affine-seed-profile-1 takes 2.315 seconds in Sim and 2.28 seconds of user CPU, compared with 3.439 and 3.40 seconds in critical-affinity-symbols-profile-1. Both use the unchanged 12-second budget, one worker, and named JIT symbols. This single pair supports a roughly 33% reduction, not a broad speedup claim. The trace is byte-identical (61e23f642f39489afa0717c767501558972f2e29419b77b088befbe8b2aff3dd), and the emitted canonical implicit programs, pure-call bodies, and complete projection plan compare equal. The new artifact has empty seeds on every affine projection stage. The critical performance gap to OMC remains open.

Prepared linear solves through conditional tensor functions

RollingWheel’s generated plan formerly lacked affine certificates for block 813 (three angle rates) and block 858 (twenty-one forces, torques, and accelerations). Both therefore entered the torn Newton solver, whose reduced Jacobian perturbs each tear variable and repeats the causal/residual sweep. OMC’s generated equations use linear systems, including the six-unknown dynamics system 731; the compilers have different state bases and block partitions, so these dimensions are not directly equivalent.

The first certificate refusal is inside the typed Frames.axesRotations owner 563: its axisRotation callees contain comparisons, conditional regions, concatenation, and numeric conversion. The constructor’s affinity walk did not cover those operations. The source formula computes a rotation matrix from the known sequence/angles and multiplies known axis/rotation coefficients by the unknown rates. Program 223, which owns the acceleration constraint at row 814, also requires concatenation, transpose, and array-patch degree bounds.

The repair extends the constructor-owned proof through both conditional regions and compact tensor operands. Unknown-dependent selectors remain nonlinear; both branch results contribute even if one is inactive at a particular point. Conversion is conservatively nonlinear in its operand. Tensor bounds cover whole ranges without per-coordinate IR expansion. This follows SPEC_0007 SOLVE-C36/C56, SPEC_0032 §2/4, and MLS §10.4, §10.6.13, and §12.4.4. No numerical tolerance, source equation, or execution program is changed.

Checked reconstruction of the newly emitted Solve model now certifies all four coupled blocks (813, 823, 858, 869) as affine. The pure-call bodies and projection plan are identical to the previous artifact; the implicit programs differ only in their derived affinity certificates. The earlier corrected row-814 dependency and all non-invalidating BLT flags remain intact. rolling-wheel/affinity-pattern-proof-1.log records the complete block census.

Two typed-call regression assertions and three tensor-operation assertions fail before their corresponding repairs. The negative controls cover unknown selectors, nonlinear coefficients, an unknown-dependent inactive branch, runtime indices/slices, and a million-element tensor. All 302 IR, 190 evaluator, 457 solver, and 128 simulation library tests pass; all-target/all-feature Clippy passes for the four crates. The fixed canary at parent 34c6ec78, working-tree digest ef383da198f992c4e6421b7dc4ecfeaff2ebde90524c490117d23b1287430aca, retains all twenty phase outcomes and bands: nine compared high, all 175 initial channels high, zero missing/skipped/nonidentifiable/deviating traces, eleven workers. Its receipt is rolling-wheel/affinity-canary-delta.json.

One clean isolated profile pair, with the same twelve-second simulation budget, shows the following changes:

Both runs have zero Sim major-fault and system-CPU deltas. The new worker SHA is f81a7c6ae2193579dbd7ecd7470612bf17a7590a8d322a14d6e345d04ea3c80f. Its trace SHA is 61e23f642f39489afa0717c767501558972f2e29419b77b088befbe8b2aff3dd; the changed arithmetic path passes the subsequent complete OMC comparison recorded above. Perf records 675 user-CPU samples without loss; trigonometry, row evaluation, and allocation remain visible costs, with incomplete native/JIT stack unwinding. The critical OMC runtime gap remains open. Combined verify quick and verify full are pending.

A second profile uses Cranelift 0.125.4’s built-in PERF_BUILDID_DIR support to emit the JIT symbol map, with no source probes. Sim takes 3.439 seconds and the trace is byte-identical to the first affinity profile. Its 672 samples have no loss; the largest named generated self costs are rumoca_jacobian_row_223 (5.65%), the directional owner 563 (4.61%), and residual program 223 (2.98%). Sine/cosine together account for 11.01% self samples. Stack unwinding still stops at generated frames, so these are not inclusive caller fractions. The exact map, perf data, and reports are retained under rolling-wheel/critical-affinity-symbols-profile-1. This points the next investigation at coefficient/Jacobian evaluation and repeated rotation work; it does not establish how much further runtime can be recovered.

OMC profile and remaining algebraic refresh overhead

Forty sequential executions of the unchanged, xtask-generated RollingWheel OMC executable confirm that solver step count does not explain the gap. The original initialization file, horizon, DASSL settings, and CSV output are retained; OMP/OpenBLAS threads are fixed to one. The executable SHA-256 is a563bc0a18b6ae9baed42ad78014c27044f94cea56cfaaae8f45ceb22793dd0a. All forty executions succeed with 1,028 DASSL steps, 1,329 ODE evaluations, 42 Jacobians, eight error-test failures, and zero convergence failures.

The whole-process wall observation includes the diagnostic driver’s polling; it is not an isolated solver timer. The simulation timer ranges from 15.868 to 21.956 ms. A 999 Hz user-CPU perf capture across these runs retains 2,502 samples with no lost samples. CSV emission accounts for 43.25% of inclusive samples and generated functionAlgebraics for 5.32%. Floating-point formatting is the largest self-time cost. The generated ODE entry invokes 44 top-level equation routines, which may themselves call other routines. Raw profiles, per-run statistics, source-artifact hashes, and commands are retained under rolling-wheel/omc-perf-repeated-1 and rolling-wheel/profile_omc_repeated.py.

Rumoca’s latest clean isolated declared-Sim measurement remains 4.283 seconds. Its profile excludes build/startup, while the OMC profile includes those phases and CSV output; their sampled percentages must not be treated as having the same denominator. A separate temporary positive-time Rumoca assignment census counts 2,224,551 calls: 2,168,913 Direct assignments and 55,638 Zero assignments. There are no TensorAffine assignments, ruling out that suspected path for this model. Instrumentation adds overhead, so its timing is not a benchmark.

The stage diagnostic identifies two refusal reasons. Coupled projection blocks without an isolatable seed for every coordinate fail the runtime seed-coverage guard. Separately, the Solve dependency certificate says projection block 868, which assigns Y114 (wheel1.rollingWheel.r_road_0[3]) to zero, invalidates earlier row 814 in coupled block 858. The runtime therefore refuses staged execution and uses the broader refresh projection. No native assignment compile or execution failures were observed. These findings are retained in rolling-wheel/critical-assignment-census-{1,4}; diagnostic probes are removed.

Inspection of the actual prepared DAE disproves an upstream BLT-order defect. Prepared expression 4023 updates component three of vector residual 3107 with expression 4021; row 814 no longer depends on algebraic coordinate 72/2, the road-height component. The lowered program preserves that exact TensorUpdate at operation 764. However, DependencyWalk::tensor_update unions the whole patch into every output and retains every original base dependency, including the overwritten component. This introduces the false reverse edge. The prepared DAE and scalar incidence are retained in rolling-wheel/ordering-prepared-dae.json and rolling-wheel/ordering-probe-2.log.

The candidate dependency derivation follows fixed index/whole-axis projections exactly, including separate AD lanes; runtime indices and slices retain their conservative dependencies. The tensor owner and operation sequence are unchanged. Three reduced vector/matrix tests fail before the fix; all four tests, including the dynamic-selector control, pass afterwards. Governing contracts are MLS §10.5/§10.6.9, SPEC_0032 §2, and SPEC_0007 SOLVE-C17/C36/C56. The separate runtime repair relies on the constructor’s complete projection block ownership instead of requiring an isolatable seed for every unknown. Missing producers still reject runtime construction; genuine backward dependencies still reject one-pass execution. Two reduced tests with derived structural artifacts fail the old dispatch requirement (coupled-stage-seeds-red-2.log); the earlier fixture omitted those artifacts and is not the decisive RED proof.

Enabling the stages exposes another runtime defect: an admitted native assignment execution error was silently retried through the interpreter. prepared-stages-libraries-3.log records the existing simulation discriminator incorrectly succeeding with an always-failing native backend. Assignment errors now propagate, restore the incoming snapshot, and do not permanently disable native execution. The discriminator now observes the assignment error, the first native failure reached by the prepared stage order. A separate regression checks restoration after a partial native write and a second attempted call.

All 297 IR, 190 evaluator, 457 solver, and 128 simulation library tests pass (prepared-stages-libraries-4.log); all-target/all-feature Clippy passes for all four crates (prepared-stages-clippy-4.log). The fixed target/msl/multibody-prepared-stages-canary passes at parent 3799c0c8, working-tree digest e5d98371f99a43da88bd65b89d3defbea7e92a14d58081eda8e0042e77f704ad. All twenty phase/simulation outcomes and bands are unchanged: nine compared models high, all 175 initial channels high, and zero skipped, missing, excluded, nonidentifiable, or deviating comparisons. Eleven workers are requested and admitted. The exact delta is rolling-wheel/prepared-stages-canary-delta.json.

Replaying the unchanged RollingWheel residual programs through the repaired artifact derivation removes the row814/Y114 edge and every algebraic reverse invalidation (prepared-stages-pattern-proof-1.log). The first isolated run takes 4.303 seconds, versus 4.283 seconds before; user CPU remains 4.24 seconds, with zero major faults during Sim. Its trace is byte-identical to the baseline, SHA-256 340e0a5bab67e9a05b160f3b50a623126edd2660863fa9cd0f25461b16bc334b. This pair establishes no speed improvement. The corrected dependency and dispatch contracts do not close the critical performance gap; the remaining prepared coupled-block execution needs profiling. The subsequent named-commit complete cohort sweep passes as recorded above; combined verify quick and verify full have not run for this change.

Selected algebraic residual native execution

RollingWheel’s selected algebraic residual callback always interpreted its source program, although selected directional callbacks already used native code. Program 223 retains 766 operations and three aggregate outputs; only one of those outputs participates in the large simultaneous block, so batching that block alone cannot remove its interpreter cost. The first responsible layer is runtime/native dispatch, not Modelica lowering.

The runtime now requests a selectable native expression for the unchanged implicit residual block. Cranelift emits each source program once and uses its entry for both selected and full evaluations. Selection uses the checked program/output coordinate, preserves complete tensor programs, and does not execute unrelated programs. Shared conditional owners retain their aggregate batch and explicitly decline selection. An admitted execution error propagates. This is target-local emission under SPEC_0032 §4 and SPEC_0007; it changes no equations, assignment ownership, solve order, state selection, or tolerances.

The isolated diagnostic pair goes from 5.144 to 4.283 seconds in declared Sim, with user CPU decreasing from 5.09 to 4.24 seconds and zero major faults during either simulation. Both complete traces have SHA-256 340e0a5bab67e9a05b160f3b50a623126edd2660863fa9cd0f25461b16bc334b. The roughly 17% reduction is one profile pair, not a stable benchmark. The OMC gap remains critical; repeated tensor assignment and directional evaluations remain visible in perf. Source hashes, worker hashes, timings, and profiles are recorded in rolling-wheel/native-residual-diagnosis.json.

Three reduced dispatch tests fail before the repair and pass afterwards. All 76 Cranelift, 456 solver, and 128 simulation library tests pass, including aggregate/sparse outputs, table context and errors, pure-call owners, shared conditional batches, explicit decline, and error propagation. All-target, all-feature Clippy passes for the three crates.

target/msl/multibody-native-residual-canary passes at parent ab36e867, working-tree digest bb06dbc1af4f6e03fcbcea70241f5220c0734b7379bcda5d7d2c95588bcd9926. All twenty phase/simulation outcomes and bands match the native-manifold canary: nine compared models high, all 175 initial channels high, and zero skipped/missing/excluded/nonidentifiable/deviating comparisons. Eleven workers were requested and admitted. Logs and the exact delta are rolling-wheel/native-residual-{libraries-2,clippy-1,canary-1}.log and rolling-wheel/native-residual-canary-delta.json. The subsequent named-commit full sweep passes with all 566 phase/simulation outcomes and bands unchanged, as recorded above. Combined verify quick and verify full have not run for this change.

Critical RollingWheel performance investigation

The 48x cached-OMC runtime gap is an open critical performance bug. An isolated worker at 8b0794d8 takes 6.094 seconds in its declared Sim phase: 6.05 seconds of user CPU, no system CPU at the counter resolution, and zero major page faults. A fresh execution of the existing OMC artifact takes 0.114 seconds including process startup and output. OMC reports 1,028 DASSL steps and 1,329 ODE calls. The earlier Rumoca census has 1,404 BDF steps; that difference does not explain the runtime gap. These diagnostic runs are not stable benchmarks.

perf and a temporary positive-time program census identify repeated tensor execution. Implicit program 223 has 766 operations and three outputs, and runs 66,589 times. Four state-manifold directional programs each run 48,880 times. The first concrete dispatch defect is in runtime preparation: manifold residual and directional blocks were never offered to the execution backend, and every projection unconditionally used the interpreter. The diagnostic receipt is rolling-wheel/critical-runtime-diagnosis.json; temporary probes are removed.

The repair offers both unchanged, source-bound manifold blocks to the existing backend once during preparation. A backend may decline compilation; an admitted native execution error propagates without an interpreter retry. This follows SPEC_0007’s checked-product ownership and SPEC_0038’s non-semantic native execution boundary. State selection, tensor owners, constraints, tolerances, and the numerical projection algorithm are unchanged. Two reduced tests fail on the old code and pass with native dispatch, interpreter-decline, state-only input, and execution-failure controls.

One isolated candidate run takes 5.144 seconds (5.09 user CPU seconds, zero major faults). Its trace is byte-identical to the baseline and to the last full run’s RollingWheel trace: 340e0a5bab67e9a05b160f3b50a623126edd2660863fa9cd0f25461b16bc334b. The remaining runtime gap is still critical. Repeated full tensor evaluations for selected algebraic residuals and sensitivities remain the next target; this improvement is not closure of the performance issue.

All 453 solver and 127 simulation library tests pass, as do all-target, all-feature Clippy checks for both crates. The fixed target/msl/multibody-native-manifold-canary passes at parent 8b0794d8, working-tree digest 6f6c0b37c0cf4b88fb8d6adc7061ca3c9a2f46d5e944857ed7c73a0be2bfb74e. All twenty phase/simulation outcomes and bands match the indexed-alias canary. Nine models compare high, all 175 initial channels are high, and there are zero skipped, missing, excluded, nonidentifiable, or deviating comparisons. Eleven workers were requested and admitted. Logs and the exact delta are rolling-wheel/native-manifold-{libraries-1,clippy-1,canary-1}.log and rolling-wheel/native-manifold-canary-delta.json. The subsequent named-commit full sweep passes with all 566 phase/simulation outcomes and bands unchanged, as recorded above; combined verify quick and verify full have not run for this change.

Fourbar indexed-alias definition repair

The original Fourbar_analytic DAE contains both position_b[i].y = k and prismatic.position_b[i] = position_b[i].y. Causal analysis interpreted the second equality as another definition of the scalar output, then discarded both definitions as duplicates. The source rows are retained in fourbar-analytic/mixed-alias-source-rows.json. OMC’s saved equation inventory assigns the three outputs 0.0, 0.2, and 0.0 during initialization (indexed-alias-omc-equations.json, equations 799–801).

The governing requirements are MLS §9.2 connection equalities and §10 array coordinates, SPEC_0007 Stage 3’s orthogonal input/output causality, and SPEC_0032 §1–2’s source-owned arrays and structural scalar views. Exact scalar/element aliases now defer their direction until independent whole definitions have been considered. A complete component definition may follow the already proved, transitively closed whole-definition graph. An externally solved algebraic cannot seed that proof, and a continuous dependency cannot become event-held. The same dependency walk supplies closure and variability; this adds no second traversal of whole-definition expression graphs.

The initial three-test reduction fails two assertions on the old code (indexed-alias-red-1.log). Five strengthened tests now cover either equality orientation, scalar alias chains, output arrays, an independently defined whole array supplying scalar outputs, incomplete coverage, duplicate scalar and component equations, cycles through the array, and a transitive continuous source. All 161 structural tests and all-target/all-feature structural Clippy pass in indexed-alias-green-4.log. Earlier fixture-construction and lint failures remain in the numbered logs.

The original DAE inspection now proves all three scalar output definitions and complete coverage of variable 476, jointSSP.prismatic.position_b. The exact before/after records and hashes are in indexed-alias-definition-delta.json. Its 829 structural reduction records remain byte-identical: differentiation still lacks consumption of complete component definitions and filters out output definitions. The original 1622/1659 structural refusal remains; this repair alone establishes neither compilation nor a new model’s trace parity. Downstream checks pass: 115 GALEC library and six integration tests, 120 Solve tests, 127 simulation tests, 572 compiler-core tests, 243 architecture tests, and seventeen size/spec gates. The log is fourbar-analytic/indexed-alias-consumers-canary-1.log.

The fixed target/msl/multibody-indexed-alias-canary passes at parent 84aa78ccb64c88647244a15b5d624732e68b2104, working-tree digest 053bb1f4e3cd33b143fe730e7905586165adde1a91bb1db1f2fea592bbb74cd4. All twenty phase and simulation outcomes and bands match multibody-state-only-manifold-canary; nine models compare high with all 175 initial channels high and zero missing, skipped, excluded, nonidentifiable, or deviating comparisons. Eleven workers were requested; the memory limiter admitted four. indexed-alias-canary-delta.json retains the comparison and artifact hashes. Tier 1 is complete; the latest full-cohort measurement remains the f485ef13 run above. No additional model pass or isolated speedup is claimed.

Previous complete measurement: singleton degree queries

target/msl/multibody-singleton-degree-full passes the local fallback gate at commit 39cb7f2384de9d15cf3d022bf16d1d2c5b1d168c, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The complete 566-model sweep takes 130.23 seconds with eleven Rumoca workers. All 152 compared models are strict-high (26.86% of 566), with seventeen reviewed exclusions, zero missing/nonidentifiable traces, and all 17,606 initial channels high. No compared trajectory channel deviates; all nine electrical counterexamples remain high. MultiBody remains 19/42 high.

There are no band changes from multibody-derivative-chain-full. PrismaticConstraint and IMC_Transformer reach structural refusals within budget. LineForceWithTwoMasses now exceeds the Solve budget instead of panicking; the separate completed DAE inspection below establishes the preflight repair, but this full attempt does not reach that final structural refusal. RollingWheelSetDriving still times out after twelve simulation seconds (12.110 seconds including call overhead), with 10.715 seconds of build time. The receipt is rolling-wheel/singleton-degree-full-delta.json under the campaign directory. The earlier twentieth high model remains a performance regression requiring repair.

The fixed 145-model performance cohort reports 2.245 median simulation speedup over OMC; the expanded 169-completion cohort reports 1.523. The medians have different populations. Neither summarizes the expensive MultiBody tail: RollingWheel takes 9.234 seconds of simulation versus OMC’s 0.134 seconds; DoublePendulum takes 1.482 versus 0.122 seconds. These exclude Rumoca build time. Combined verify quick last passed before six subsequent implementation changes; verify full has not run. No PR or baseline promotion is claimed.

RollingWheel runtime census and state-only projection

The MSL worker uses the in-process FMI Model Exchange host, Diffsol BDF for this model, Cranelift where execution is admitted, and prepared interpretation for remaining work. These measurements do not execute an exported C FMU. The current Solve artifact retains twelve scalar states, 899 algebraic slots, and four state constraints; OMC selects eight states. A temporary numerical census records 1,404 BDF steps, 27 resets, 336 linear setups, and 2,324 nonlinear iterations through 4 seconds. Accepted orders 1 through 5 have counts 54/255/242/339/514. OMC’s generated DASSL executable takes 1,028 steps, 1,329 ODE calls, 42 Jacobian evaluations, and zero events. A permanently low BDF order therefore does not explain the large runtime gap.

The refresh census records 6,843 full-algebraic refreshes taking 6.748 seconds, 3,670 derivative refreshes taking 2.279 seconds, and 2,512 root refreshes taking 0.189 seconds. These counters include preparation, initialization, and simulation; they are not exclusive declared-Sim samples. Both instrumented trajectories are byte-identical to the complete sweep’s RollingWheel trace. The temporary probes are removed; exact patches, binary/source hashes, logs, and scope are retained in rolling-wheel/rolling-wheel-runtime-census.json. The first OMC executable invocation lacked its relative sparsity assets; that setup failure and the corrected working-directory invocation are both retained.

The concrete ordering defect is in the FMI kernel’s state projection: it reconstructs all algebraics before evaluating the retained state-only manifold. Solve lowering already rejects algebraic/derivative dependencies in that manifold (lower.rs::manifold_state_slots, SPEC_0007 structural/Solve boundary). The reduced regression projects an off-manifold state from -1 to 4 before observing a*a=x; the old code fails trying to find a real a at -1. Its negative control keeps an invalid post-projection observation as an error. The candidate passes only the state prefix to the unchanged residual/JVP projection, preserving tensor owners, numerical tolerances, and the accepted step’s existing restart policy. An invalid manifold that reads an algebraic warm start is refused by the checked evaluator and leaves the state unchanged. All 451 solver tests, all-target/all-feature solver Clippy, formatting, 243 architecture tests, and seventeen size/spec gates pass. The positive regression first failed on the old premature algebraic solve; the invalid-output control already passed. Logs are rolling-wheel/state-only-manifold-red-1.log, state-only-manifold-green-2.log, and state-only-manifold-canary-1.log. The fixed twenty-model target/msl/multibody-state-only-manifold-canary has no phase or band changes: nine comparisons and all 175 initial channels high, zero missing/skipped/excluded/nonidentifiable traces. It ran at parent 39cb7f23 with dirty digest a3031f39be8bb4047180ee0e1142d4a5c9a43b65010c3b9b861017d54a8008a0. Eleven workers were requested, but the memory limiter admitted one; this is not an eleven-worker timing comparison. The delta is retained in rolling-wheel/state-only-manifold-canary-delta.json. The completed full sweep at f485ef13 is recorded above.

Previous complete measurement: supplied derivative chains

target/msl/multibody-derivative-chain-full passes the local fallback gate at commit 7fa18fa1c93b4b65c588e98da6509aaf794dc03d, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The complete 566-model sweep takes 124.74 seconds with eleven Rumoca workers. All 152 compared models are strict-high (26.86% of 566), with seventeen reviewed exclusions, zero missing/nonidentifiable traces, and all 17,606 initial channels high. No compared trajectory channel deviates; all nine electrical counterexamples retain their high bands.

MultiBody measures 19/42 high. RollingWheelSetDriving initializes but again exhausts its twelve-second simulation budget (12.113 seconds including call overhead). It is the only band loss from multibody-shared-call-full. PrismaticConstraint and IMC_Transformer change from structural refusals to Solve timeouts. LineForceWithTwoMasses reaches a compiler panic instead of its previous Solve timeout: holonomic preflight proves a causal algebraic definition. The panic is the immediate triage priority; the wheel-set performance loss also remains open. A passing fallback gate does not discharge either defect. fourbar-analytic/derivative-chain-full-delta.json retains all four phase deltas, the band delta, electrical checks, and result/comparison/band artifact hashes.

Fourbar_analytic still reports 1622/1659 structural matches. The combined verify quick passed before the last four implementation changes; focused checks and the fixed canary pass as recorded below. verify full has not run. No PR, release, or baseline promotion is claimed.

Previous complete measurement: restored wheel-set completion

target/msl/multibody-shared-call-full passes the local fallback gate at commit 51459c19c3aa59db8925b3a0d8eb40846d666d67, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The complete 566-model sweep takes 140.11 seconds with eleven workers. All 153 compared models are strict-high (27.03% of 566), with seventeen reviewed exclusions and zero missing or nonidentifiable traces. All 18,498 initial channels are high and no trajectory channel deviates. All nine named electrical counterexamples retain their high bands.

MultiBody returns to 20/42 high. RollingWheelSetDriving completes in 11.304 seconds, with all 892 trajectory and initial channels high. Its worst channel’s bounded normalized L1 score remains 5.061492181479045e-6. Build takes 11.199 seconds, including 7.809 seconds of Solve lowering. RollingWheel remains high at 8.682 seconds of simulation. The wheel-set restoration is the only phase/simulation or band change from multibody-parameter-branch-full; the receipt and artifact digests are in rolling-wheel/shared-call-full-delta.json. The solver budget remains twelve seconds. This completes the shared-interface equality fix’s cohort check.

Thirty-five of 42 MultiBody examples reach the DAE boundary. The remaining 22 without high parity comprise thirteen timeouts (two Flatten, ten Solve, one simulation), four structural refusals, and five earlier producer errors (three record derivative-specialization refusals, one function-shape refusal, one array-valued fixed attribute refusal). Fourbar_analytic still reports 1622/1659 structural matches; its OMC derivative chain is the next focus. At that checkpoint, the combined verify quick passed before the preceding three implementation changes; the focused suites, canary, and full MSL checks pass as recorded below. verify full has not run. No PR, release, or baseline promotion is claimed.

Previous complete measurement: wheel-set timeout

target/msl/multibody-parameter-branch-full passes the local fallback gate at commit c89fde703f82afc323d9f38bdf1e8f316452d723, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The complete 566-model sweep takes 123.99 seconds with eleven workers. All 152 compared models are strict-high (26.86% of 566), with seventeen reviewed exclusions and zero missing or nonidentifiable traces. Every compared initial channel is high and no trajectory channel deviates. All nine electrical counterexamples retain their high bands.

MultiBody returns to 19/42 high because RollingWheelSetDriving times out after successful initialization. Its enclosing simulation call takes 12.130 seconds and reports timeout after 12.000s; build takes 11.051 seconds, including 7.646 seconds of Solve lowering. This is the only phase/simulation or band change from multibody-native-singleton-full, retained in fourbar-analytic/parameter-branch-full-delta.json. The earlier twentieth high model is a real trace comparison but not yet reliable completion. The passing fallback gate does not excuse this loss. Performance work takes priority over further capability changes; no timeout or comparator change is proposed.

A declared-Sim perf diagnostic on RollingWheelSetDriving completes in 11.875 seconds, capturing 2,342 samples with zero lost. This diagnostic does not replace the cohort timeout. Of 163 memcmp samples, captured direct return addresses identify 95 dependency-slice comparisons, 28 input-type comparisons, 23 output-interface comparisons, and seventeen comparisons inside native call_cells. The original DWARF caller unwind fails. The bounded replacement reads the first user-stack word of leaf libc memcmp, using the ELF load bias derived independently from a resolved Rust symbol; it makes no recursive-unwinding claim. Evidence is under rolling-wheel/wheel-set-runtime-profile, including memcmp-return-addresses-corrected.json. Source inspection and disassembly show repeated value comparisons of immutable interface slices already shared by Arc. A sharing-aware equality fast path is the next hypothesis.

Shared call-interface equality

The owner-table equality query now recognizes identical immutable Arc slice allocations before comparing their values. The helper requires Eq, whose reflexivity justifies the shortcut; a merely PartialEq element could not justify it. Independently reconstructed interfaces still undergo the same full value comparison, and primal/directional owner checks remain required. This is a read-only data-integrity optimization under SPEC_0029 §3, not a new semantic identity, wire format, or execution policy.

The focused RED visits all three elements when comparing shared storage (rolling-wheel/shared-call-red-2.log). After the fix that count is zero; independent equal and unequal slices still compare all three values. Wire replay retains value equality and rejects altered input, output, and dependency slices in both primal and directional interfaces. All 293 Solve IR tests, 190 evaluator tests, 72 native-backend tests, and 447 solver tests pass. All-target/all-feature IR Clippy, 243 architecture tests, seventeen size/spec gates, formatting, and whitespace checks pass (rolling-wheel/shared-call-green-1.log, shared-call-gates-build-1.log).

The matched declared-Sim diagnostic completes in 11.474 seconds versus 11.875 before. memcmp falls from 6.96% to 1.11% of samples; the new capture has 2,250 samples with zero lost. The emitted traces are byte-identical, SHA-256 3dc2333f4b9d6d69bead292633f3304aa57204f430040d50c00fb62753006a5b. Build time rises from 11.613 to 12.307 seconds, including Solve lowering from 7.371 to 8.051; the change targets repeated runtime comparisons. The single pair proves the targeted cost was removed without changing this trace, not a stable speedup or restored cohort coverage. Receipt: rolling-wheel/shared-call-profile-pair.json.

The fixed twenty-model target/msl/multibody-shared-call-canary has no phase or band changes from multibody-parameter-branch-canary. Its nine compared models and all 175 initial channels remain high, with zero skipped, missing, excluded, or nonidentifiable traces. This is Tier 1 evidence only, captured on f3db93c41e049e5d9eb16720cf36bf4f68c0cc6f with candidate worktree digest 80dfdfee5fdfc0c2453ca7e42c7f1168148124844e8d2cc68aaf294a5083f04b; rolling-wheel/shared-call-canary-delta.json retains the receipt. The next complete sweep must decide whether the wheel-set completion loss is closed.

Previous complete measurement: first RollingWheelSetDriving high trace

target/msl/multibody-native-singleton-full passes the local fallback gate at commit 27bc2c8d711703fc679b492a26e542486bb19019, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The complete 566-model sweep takes 138.19 seconds with eleven workers. All 153 compared models are strict-high (27.03% of 566), with seventeen reviewed exclusions and zero missing or nonidentifiable traces. Every previously high model retains its band, including all nine original electrical counterexamples. All 18,498 initial channels are high and no trajectory channel deviates.

MultiBody reaches 20/42 high, with twenty compared and no trace boundaries. RollingWheelSetDriving advances from a twelve-second simulation timeout to high parity on all 892 trajectory and initial channels. Its worst channel’s bounded normalized L1 score is 5.061492181479045e-6. The enclosing simulation call takes 12.050 seconds, including its own initialization work; the solver budget remains twelve seconds and the parent watchdog remains fourteen. Build takes 10.467 seconds, including 7.194 seconds of Solve lowering, and the separate initialization check takes 0.142 seconds. This is a measured high result with little runtime margin, not proof of reliable completion under every host load. RollingWheel retains all 184 high channels at 9.261 seconds of simulation.

IMC_Transformer returns from a Solve timeout to its structural refusal. These are the only two phase/simulation changes relative to multibody-quick-after-architecture-full; RollingWheelSetDriving is the only band change. The retained receipt and artifact hashes are in rolling-wheel/native-singleton-full-delta.json. No tolerance, baseline, timeout, target-list, or exclusion changes were made. verify quick passed before this implementation, and its focused tests and gates pass as recorded below. verify full has not run. Twenty-two MultiBody examples still lack high parity; Fourbar_analytic’s structural refusal is the next investigation.

LineForceWithTwoMasses: retained-value preflight

The full derivative-chain sweep exposed a panic while restoring a previously retained manifold after direct state demotion. A bounded replay identifies body1.w_a, RHS expression 7059, as differentiable but not reconstructible as a state-only value. Reconstruction reaches jointUPS.axisLength without a causal value anchor and hits an expect whose alleged preflight never ran. This is a structural proof/consumer mismatch, not a Modelica source error or solver issue. The earlier candidate for the same state, RHS 5693, does have a value proof.

OMC’s flat, optimiser, and backend XML captures under line-force/omc-stages all succeed. Its flat source defines body1.w_a = angularVelocity2(body1.frame_a.R); the function returns the record’s w field. Its backend represents axisLength as a dummy state, retains the geometric square-root definition, and issues its first and second derivatives. Rumoca’s panic is earlier than numerical execution; these captures do not establish a Rumoca trace comparison.

The reduced checked-DAE test returns a whole vector from a pure function whose other argument is a matrix. Differentiation can follow the returned argument, but retaining the complete call needs the matrix’s exact value. The RED reproduces the same panic (line-force/manifold-preflight-red-1.log). Before reconstructing a demotion, structural analysis now proves an exact RHS value if a surviving manifold references that state. The value must use remaining state/invariant anchors. A missing proof rejects this candidate through the existing WouldInvalidateManifold outcome, allowing other candidates to be considered. A removed lifted constraint requires no retained-value substitution. Tensor arguments, source equations, assertions, and the final manifold check remain under their existing owners (SPEC_0007 structural scope / STRUCT-T03).

The negative case now rejects without panic; a control with an exact whole-matrix definition reconstructs a state-only manifold, and demotion without a retained constraint remains available. All 156 structural tests, 572 compiler-core tests, 243 architecture tests, seventeen size/spec gates, structural Clippy, formatting, and whitespace checks pass. No temporary probes remain. The original DAE replay now completes with a typed structural refusal: its best intermediate residue is one equation/unknown pair, but no complete reduction is established. All 1,143 reduction records, including two manifold rejections, are retained with the source DAE digest in line-force/manifold-preflight-inspection.json.

The fixed twenty-model target/msl/multibody-manifold-preflight-canary has no phase or band changes from multibody-derivative-chain-canary. Its nine compared models and 175 initial channels are high, with zero skipped, excluded, missing, or nonidentifiable traces. This is Tier 1 evidence on parent c7d85fc3b9b068f249d7f2972895c01559932e7b, candidate worktree digest 9be3cae1653ab969bc61f1a89a7dea73b0bee0567ac6638d7538bc096fb0f3f5; receipt line-force/manifold-preflight-canary-delta.json. The latest cohort number remains the named derivative-chain sweep above. Wheel-set performance remains open before further capability expansion.

Wheel-set follow-up profile after the preflight repair

One declared-Sim perf diagnostic at d3c414d8 completes in 11.670 seconds, with 2,303 samples and zero lost. Build takes 9.824 seconds, including 6.957 seconds of Solve lowering. The trace is byte-identical to the earlier high wheel-set trace, SHA-256 3dc2333f4b9d6d69bead292633f3304aa57204f430040d50c00fb62753006a5b. This single-worker diagnostic does not replace the eleven-worker cohort timeout. The profiler is attached only during the worker’s declared Sim phase, with the same twelve-second solver budget (rolling-wheel/manifold-profile-receipt.json).

The main leaf costs remain prepared-row interpretation (9.08%), sparse LU (5.17%), native call payload conversion (2.74%), projection-affinity lookup (2.52%), assignment-certificate queries (2.26%), and row/output lookup (1.65%). The earlier shared-interface comparison fix remains effective: memcmp is 1.13%. Caller unwinding is still almost entirely absent, so these are leaf samples, not inclusive cost attribution. Next hypotheses are preparing repeated immutable projection queries once and examining aggregate projection execution. The existing native singleton path deliberately excludes multi-output tensor rows; any extension must preserve one tensor program owner and its checked assignment semantics. No runtime optimization is claimed from this capture.

Untorn singleton projection before degree queries

For a block without a tearing, both degree classifications attempt the same singleton assignment before residual projection. Runtime now performs that common attempt first, then asks for the degree certificate only if it declines. It does not repeat a declined assignment. A block with a tearing retains the previous affine/torn/assignment ordering and its branch selection. The residual Newton path is factored into its own helper without changing its arithmetic. This is a control-flow optimization over existing constructor-issued facts (SPEC_0029 §5 and SPEC_0036/SPEC_0043 §6a); tensor source programs, assertion execution, assignment acceptance, and rollback semantics retain their owners.

The focused RED records one redundant degree query for an exact singleton; the candidate records zero with either an affine or unproved classification. A torn-singleton control still requests degree selection, and an inexact candidate still reaches residual Newton. All 448 solver tests, all-target/ all-feature solver Clippy, 243 architecture tests, seventeen size/spec gates, formatting, and whitespace checks pass (rolling-wheel/singleton-degree-red-1.log, singleton-degree-green-2.log, singleton-degree-canary-1.log).

The matched declared-Sim wheel-set profile takes 11.183 seconds versus 11.670. The degree-query hotspot falls from 2.52% of samples to zero recorded samples; 2,214 samples are captured with zero lost. The complete trace remains byte-identical, SHA-256 3dc2333f4b9d6d69bead292633f3304aa57204f430040d50c00fb62753006a5b. Build takes 11.098 seconds, including 7.008 seconds of Solve lowering. The single diagnostic pair proves the targeted cost and unchanged trace, not a stable speedup or restored cohort completion. Its receipt and exact candidate patch digest are in rolling-wheel/singleton-degree-profile-pair.json.

The fixed twenty-model target/msl/multibody-singleton-degree-canary has no phase or band deltas from multibody-manifold-preflight-canary. Its nine compared models and 175 initial channels remain high, with zero skipped, excluded, missing, or nonidentifiable traces. This Tier 1 result is on parent 45b3507e00b0b073573864eeffd8a0ea64d40e6d, candidate worktree digest e5ce5b73a334ab799b53e6003fdd031e353d60e285e0aaf268529c25b75ebb77; receipt rolling-wheel/singleton-degree-canary-delta.json. The following full cohort sweep must decide whether the wheel-set timeout is closed.

Fourbar_analytic: supplied derivative chains

MLS §12.7.1 permits a higher-order annotation only in the differentiation chain that issued its arguments. The candidate follows checked predecessor links in source priority order and records each original argument together with its derivative order. Each link appends only derivatives of the previous link’s last tangent group. Reconstruction consumes this same compact plan; whole tensor arguments remain whole. Governing contracts are SPEC_0007’s structural stage, SPEC_0032, and SPEC_0036/SPEC_0043 §10.

The reduced oscillator has der(q)=v, der(v)=-q, a position function with the same max-like branch as the MSL joint, and both kinematic derivatives. Its first-order control passes, while requesting acceleration fails at 4/5 structural matches (fourbar-analytic/derivative-chain-red-1.log). OMC accepts both branches and the first-order control. Additional quadratic-function and direct-call controls verify the full argument history; all seven OMC cases match their analytic trajectories within 4.40e-8 absolute error. The two analytic-comparison.json receipts are under derivative-chain-omc and derivative-chain-quadratic-omc.

Extending selection alone is insufficient: holonomic value materialization and its first-derivative proof also looked through the primal function body. Value reconstruction now retains the checked pure call after proving every argument reconstructible from state/invariant values. Its supplied first derivative is checked through the same selected argument plan. This preserves the original assertion behavior and avoids requiring a parameter-only guard proof inside the supplied function. The reduced second-order system retains two dependent states with two manifold constraints, leaving two independent states; both BDF and RK traces agree with the analytic oscillator.

Six focused tests cover both branches, a nonlinear second derivative, whole tensor argument/result shapes, a direct call lacking the higher-order context, and both valid and invalid assertion domains. The quadratic direct-call control supplies q_d=2*v; incorrectly assuming this is der(q) would change its derivative by 4*v*v. It retains the structural refusal. All 154 structural tests and the then-571 compiler-core tests pass, followed by the expanded six-case suite and thirty derivative regressions. Structural and compiler-core Clippy, 243 architecture tests, seventeen size/spec gates, formatting, and whitespace checks pass. Logs include derivative-chain-gates-build-{1,2}.log, derivative-chain-green-6.log, and derivative-chain-final-checks-2.log under fourbar-analytic.

The fixed twenty-model target/msl/multibody-derivative-chain-canary has no phase or band changes from multibody-shared-call-canary. Its nine compared models and all 175 initial channels remain high, with zero skipped, missing, excluded, or nonidentifiable traces. This Tier 1 capture uses parent 5d0f62d147caa94e04befb9532f437b0b16cb9de and candidate worktree digest 7a0bca0656d146c6225b43f0262d9039b35092b9b6966bda0ed399af6cd1b641; fourbar-analytic/derivative-chain-canary-delta.json retains the exact delta.

The originating target/msl/multibody-derivative-chain-fourbar attempt still reports 1622/1659 structural matches, with 7.872 seconds of build time and no simulation. Its gate correctly fails as parity unmeasured. The current inspector again emits the same 829 reduction records (reduction-after-derivative-chain-1.log). A bounded generic rejection probe, removed after capture, identifies an earlier obstruction: the joint’s prismatic.position_b input is driven by three Constant-block outputs. Flatten emits three indexed connection equations (source-stages/ir-flat.mo lines 991–993), but variable 476 has no whole-vector causal definition. The differentiation proof stops on this vector before reaching the supplied derivative chain. rejection-probe-2.log retains that evidence. The next investigation is preservation of the source aggregate connection through Flatten and ToDAE; no coverage gain is attributed to this reduced fix yet.

Fourbar_analytic: differentiating a parameter-selected position

The existing omc-structure inspector reproduces the original 1622/1659 structural refusal. An observed reduction of the exact source DAE shows why that final diagnostic alone is insufficient: direct demotion reaches six unmatched equations, while the later pristine holonomic path reaches eight. The six-equation frontier names orientation/position equations and leaves jointSSP.totalPower, body accelerations, and b2.body.frame_a.t[1] unmatched. The reducer ultimately returns its preserved earlier refusal. Logs are fourbar-analytic/structure-1.log and fourbar-analytic/reduction-1.log.

Rebuilding the structural inspector after the parameter-branch change and replaying the same checked source DAE produces exactly the same 829 reduction records. Both lanes reach a minimum residue of twelve (six unmatched equation/unknown pairs); the later holonomic lane finishes at sixteen. The record-stream SHA-256 is 0936b7735c12449841f9fb2a55e5f1092ca8aea833ed3dcbab0cc62b2ed6434c before and after. fourbar-analytic/parameter-branch-reduction-delta.json binds this result to the current capture. The reduced parameter case is repaired, but this evidence gives it no credit for advancing Fourbar_analytic.

OMC stage captures include flattened Modelica, flat/optimiser/backend XML, and transformation-debugger equations in fourbar-analytic/omc-stages. OMC selects exactly j1.phi and j1.w as states, with one linear torn system (one iteration variable and thirty inner variables), no nonlinear torn system. Its equations explicitly propagate first and second derivatives through the prismatic joint’s quadratic position solution.

The source joint computes distance = -k1 + (if positiveBranch then k2 else -k2), where positiveBranch is a Boolean parameter solved during initialization. Structural differentiation had no conditional-expression case unless the guard could already be selected as a literal. Parameter bindings deliberately are not literals: initialization and parameter updates must still choose the branch.

The reduced ParameterBranchKinematics oscillator preserves two forced states, observes a parameter-selected signed position, and requests both kinematic derivatives. The unconditional alias control passes. Both a bound Boolean parameter and a fixed=false initialized Boolean fail structural analysis at 4/5 matches, leaving the position equation and acceleration unmatched (fourbar-analytic/parameter-branch-red-2.log). OMC accepts the literal, initialized-positive, and initialized-negative sources; all five channels match the analytic oscillator within 9.44e-9 absolute error (fourbar-analytic/parameter-branch-omc/analytic-comparison.json).

The candidate derives guard invariance from parameter coordinates and their checked operations/call-argument substitutions, requires every branch value to satisfy the relevant differentiation/materialization proof, and retains the original conditions and branch order around the differentiated values. Tensor branches retain their whole shape, including zero derivatives. Governing requirements are MLS §3.6.5, §3.8.3, §8.6, SPEC_0007’s structural stage contract, SPEC_0032, and SPEC_0036. All six focused tests pass, including function substitution, both tensor branches, and rejection of time/state guards from this parameter-only proof. The tensor test audits the prepared DAE: the position and both derivative conditionals retain vector shape [2]. All 154 structural and 566 compiler-core tests pass (fourbar-analytic/parameter-branch-gates-1.log). All-target, all-feature structural Clippy, 243 architecture tests, and seventeen size/spec gates pass after replacing a manual even-index test with is_multiple_of (fourbar-analytic/parameter-branch-gates-build-2.log). Formatting and whitespace checks pass.

The normal-budget target/msl/multibody-parameter-branch-fourbar attempt still returns 1622/1659 structural matches, in 8.389 seconds of build time. It produces no simulation or comparison: the focused gate correctly reports parity unmeasured. This reduced repair does not close the original model. The fixed multibody-parameter-branch-canary passes with all twenty phase/simulation and band rows unchanged from multibody-native-singleton-canary. All nine compared models and 175 initial channels remain high, with zero deviating, skipped, missing, excluded, or nonidentifiable comparisons. It runs at b102b3f7, dirty-tree digest 93d23c8089bdbfe881e2f3c80952ec46d792d48a9955f051fc6a39558860c837; the receipt is fourbar-analytic/parameter-branch-canary-delta.json. Tier 1 is complete. A named-commit cohort sweep and continued equation-level investigation of Fourbar_analytic remain next.

The DAE JSON capture succeeds, but readable DAE rendering refuses a function assertion (unsupported-feature:dae-modelica-function-statement:assertion). The flattened text and exact DAE JSON remain available for this investigation; that rendering limitation is separate from structural compilation.

Previous complete measurement: quick-suite success

The MSL step of verify quick --early-exit at commit be391b3f25e216616af9167a16c48f5e95c6eb1d, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35, passes the local fallback gate in 296.88 seconds with the quick suite’s four model workers. Its result, comparison, and band table are preserved in target/msl/multibody-quick-after-architecture-full. All 152 compared models are strict-high (26.86% of 566), with seventeen reviewed exclusions and zero missing or nonidentifiable traces. All 17,606 initial channels are high; no trajectory channel deviates. Every high model from multibody-function-scope-index-full retains its band, including all nine original electrical counterexamples.

MultiBody is 19/42 high, with nineteen compared and no trace boundaries. RollingWheel completes with all 184 channels high and 9.656 seconds of simulation including initialization. Its earlier twelve-second timeout remains recorded below: this run follows an architecture repair, not a demonstrated runtime performance fix. IMC_Transformer still changes from a structural refusal to a Solve timeout relative to multibody-function-scope-index-full; the full delta is rolling-wheel/verify-quick-architecture-msl-delta.json.

The combined verify quick --early-exit suite passes in 1396.72 seconds: workspace lint, MSL parity, all 28 pinned corpus rows (89.3 seconds), 243 architecture-hardening tests, seventeen size/spec gates, all 7,516 nextest tests across 146 binaries, and workspace doctests. Existing ignored doctest examples remain ignored. The log is .git/multibody-campaign/verify-quick-after-architecture-1.log; the timing receipt is rolling-wheel/verify-quick-after-architecture-passed.json. This validates the implementation at be391b3f; the subsequent 6ba7bb80 commit only records evidence. verify full has not run. Further explicit MSL runs retain the approved eleven simulation workers.

Native execution of singleton projection assignments

A fresh RollingWheel perf capture attaches at the worker’s declared Sim phase, after build and the separate initialization check. The pre-change rolling-wheel/quick-rolling-runtime-profile capture records 1,799 samples, zero lost, and 9.115 seconds of simulation. Prepared row interpretation has 12.95% self samples. Caller unwinding is incomplete, so this is not an inclusive cost attribution to any particular projection caller.

Source inspection identifies an interpreter-only consumer in RefreshProjectionModel::eval_implicit_target_value. The candidate hands its existing constructor-derived exact assignment to the optional native backend and caches compilation by immutable program/target identity. Only source programs with one stored scalar output are admitted; aggregate programs keep their shared owner and existing execution path. The compiler’s assignment materializer preserves independent tiny offsets and singularity guards (SPEC_0036 and SPEC_0043 §6a, SPEC_0032). No equation, state selection, tolerance, or model-name policy changes.

The focused regression first records two failures: no native execution and an injected native error hidden by the interpreter-only path. All six new tests pass, covering reuse, compilation refusal caching, error propagation, aggregate retention, tiny roots, and identical singular-coefficient declines. The simulator’s existing injected-failure test initially expected successful interpreter completion; it now requires the native projection error while retaining its zero-success and nonzero-failure assertions. All 447 solver and 127 simulator library tests pass. Logs are rolling-wheel/native-singleton-red-2.log, rolling-wheel/native-singleton-tests-build-1.log, and rolling-wheel/native-singleton-tests-build-2.log.

The matching post-change capture, rolling-wheel/native-singleton-runtime-profile, completes simulation in 8.677 seconds against the same twelve-second budget. Its complete trace JSON is byte-identical to the pre-change trace (f8df74c085fa7a397df10ebdf9754cc48bd2bf34776aa81ebb85b9e5ac23e31c). This single pair observes a 4.8% reduction; it does not establish a stable speedup or replace a cohort comparison. Prepared interpretation has 11.39% self samples. Build and separate initialization check change from 4.146 / 0.471 seconds to 3.841 / 0.451 seconds. The timing receipt is rolling-wheel/native-singleton-runtime-timing-delta.json.

The fixed target/msl/multibody-native-singleton-canary passes in 9.49 seconds with eleven workers at HEAD 6ba7bb80, working-tree digest 7c13adb6fd70e33d052159b2b017ed0def8711e7f1d85defd32be26d33a4d81a. All twenty phase/simulation and band outcomes match multibody-function-scope-index-canary. All nine compared models and 175 initial channels remain high, with no deviating, missing, skipped, excluded, or nonidentifiable comparisons. The retained receipt is rolling-wheel/native-singleton-canary-delta.json. Tier 1 is complete; the next full cohort measurement remains pending. All-target, all-feature Clippy for solver/simulator, 243 architecture-hardening tests, and seventeen size/spec gates pass (rolling-wheel/native-singleton-gates-1.log and rolling-wheel/native-singleton-size-gates-1.log). The first combined command named the nonexistent code_size_test target after its successful Clippy and architecture steps; the subsequent command runs the actual suite_gates target. Formatting and whitespace checks pass.

Previous complete measurement: quick-suite gate failures

The MSL step of verify quick --early-exit at commit 03865c4e9e2fb3e4c6cae86b1b9b233cd5237e91, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35, compares 151 models, all strict-high (26.68% of 566), with seventeen reviewed exclusions and zero missing or nonidentifiable traces. All 17,422 initial channels are high, and no trajectory channel deviates. Its result, comparison, and band table are preserved in target/msl/multibody-quick-after-scalar-contact-full; the delta is rolling-wheel/verify-quick-msl-delta.json.

MultiBody falls to 18/42 high because RollingWheel exceeds the unchanged twelve-second simulation budget, after successful initialization. This is an execution regression, with no observed deviating trace. PrismaticConstraint and IMC_Transformer also change from structural refusals to Solve timeouts. This invocation used the quick suite’s automatic four-worker schedule, while the preceding explicit full run used eleven. The timeout is retained as a failure; neither the passing fallback gate nor earlier successful measurements establish reliable completion. All nine original electrical counterexamples retain high parity.

The combined quick suite passes lint, MSL parity, and the pinned external corpus, then fails its architecture gate: Solve has 274 runtime totality assertions against the unchanged ceiling of 272. Workspace tests do not run after this early exit. Neither verify quick nor verify full is green. The repair selects binary tensor builtin kinds and binary scalar operators once in Solve lowering, then consumes those narrowed values. This removes two impossible-case assertions without changing operand checks, operation ordering, tensor ownership, or the gate. All 120 Solve library tests and 243 architecture-hardening tests pass. The next gate exposes oversized specs: SPEC_0007 has 2680 words; SPEC_0036 has 2810 words and 367 lines. Their detailed target-product, continuous-refresh, relation, and initialization requirements move verbatim into the existing SPEC_0040/0043 catalogs, retaining mandatory parent links and every requirement. No status or budget changes. The parents now contain 2403/2414 words and 328/325 lines. All 243 architecture-hardening and seventeen size/spec gates pass. rolling-wheel/spec-catalog-move-receipt.json verifies each moved passage byte-for-byte against the prior commit. RollingWheel performance remains the next investigation before MultiBody breadth work resumes.

Previous complete measurement: function scope indexing

target/msl/multibody-function-scope-index-full passes the local fallback MSL gate at commit a2bf9abc7ea12a0c16a2674e22626d9abfd03c92, digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The 566-model run finishes in 136.49 seconds with eleven model workers. All 152 compared models remain strict-high (26.86% of 566), with seventeen existing reviewed exclusions and zero missing or nonidentifiable traces. All 17,606 initial channels are high; no trajectory channel deviates. Every previously high model, including the nine original electrical counterexamples, retains its band.

All four Solve completion regressions from the scalar-contact sweep recover. DCPM_Drive completes again under its unchanged reviewed exclusion, with 9.185 seconds of Solve lowering and 9.383 seconds of build time. PrismaticConstraint, GearConstraint, and GenerationOfFMUs again report their structural refusals within budget. Raw simulation and initialization completions return to 169 and 183. Solve completion reaches 258: Fourbar_analytic now reports a structural refusal, 1622/1659, instead of timing out. This gives a concrete next diagnostic, not a new successful simulation.

MultiBody remains 19/42 high, with nineteen compared, all 8,477 trajectory and initial channels high, and no skipped, missing, excluded, or nonidentifiable MultiBody traces. DAE completion remains 35/42. RollingWheel reports 9.264 seconds of simulation including initialization. Relative to the pre-contact multibody-selected-jvp-full, every simulation band and execution outcome is retained; only Fourbar_analytic changes its refusal classification. The full delta and hashes are in rolling-wheel/function-scope-index-full-delta.json. Combined verify quick and verify full success and complete MultiBody coverage are not established.

Previous complete measurement: scalar contact reconstruction

target/msl/multibody-scalar-contact-full passes the local fallback-baseline MSL gate at commit adc5d4b2ac965d1b1778647ad78bc26dc16f5a6a, digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The full 566-model comparison finishes in 125.72 seconds with eleven model workers. All 152 compared models remain strict-high (26.86% of 566), including all nine original electrical counterexamples. There are 16 existing reviewed exclusions and zero missing or nonidentifiable traces. All 17,606 initial channels are high; trajectory channels comprise 17,589 high and seventeen minor, with none deviating or severe.

MultiBody remains 19/42 high, with nineteen compared and all 8,477 trajectory and initial channels high. There are no MultiBody trace boundaries; DAE completion remains 35/42. No additional example closes in this run. The repaired contact-coordinate core regression is independently green.

Four models lose completion of the Solve phase: PrismaticConstraint, GearConstraint, GenerationOfFMUs, and DCPM_Drive. The first three previously reported structural refusals; DCPM_Drive previously completed simulation under an existing reviewed exclusion and now hits the ten-second Solve budget. This lowers raw simulation completions from 169 to 168, initial completions from 183 to 182, and Solve completion from 257 to 253. The exclusion policy and all model bands are unchanged. The passing baseline gate does not excuse these execution regressions; performance repair takes priority over further capability work. The full delta and hashes are in rolling-wheel/scalar-contact-full-delta.json.

A bounded standalone DCPM_Drive perf diagnostic captures 5,489 samples with zero lost samples. It completes simulation, but Solve lowering takes 11.436 seconds and total build takes 12.168 seconds, confirming work beyond the cohort’s unchanged phase budget. Function reconstruction accounts for 16.25% of self samples, causal-definition derivation 7.58%, and scalar affine recognition 2.23%. Source inspection finds that orphaned-expression recovery scans the full expression arena once per function. Indexing those existing function scopes once is the next hypothesis; no repair is claimed yet. This diagnostic completion does not replace the full-run timeout.

Function scope indexing

The performance repair extends the existing function-expression inventory pass with per-function scope lists. Orphan recovery visits each function’s existing branded expression identities in the same ascending source order; the final completeness check remains intact. This removes the repeated whole-arena scan without changing function statements, definitions, or construction order. All 154 structural library tests pass.

The matched standalone DCPM_Drive diagnostic reduces Solve lowering from 11.436 to 7.804 seconds and build time from 12.168 to 8.302 seconds. The new perf capture has 9,153 samples and zero lost samples. These are diagnostic timings, not a statistical benchmark. Separate before/after artifact runs produce byte-identical DAE, Solve IR, and simulation traces: the Solve SHA-256 is 095f50d80afa1cf221f3362fa4195d3b607b4899715bad02e9f6d7e4d926bcb3 and the trace SHA-256 is b80f5af6dc932435adfc419704b75e0b5553f38cd520cedc316d78c8d29209ad. rolling-wheel/function-scope-index-equivalence.json records all hashes. Affected all-target/all-feature Clippy checks and formatting pass. The fixed target/msl/multibody-function-scope-index-canary passes with all twenty phase, simulation, initial-condition, and band outcomes unchanged. Nine models compare high, all 175 initial channels are high, and skipped, missing, excluded, nonidentifiable, and deviating counts are zero. At HEAD 09e66c8be6c5df71fbf964564516ff15eb678d0c, its working-tree digest is 94c4aafbbe07bf098a355f7069b749bfae1541297682197261500d1d2024eeca. rolling-wheel/function-scope-index-canary-delta.json records the delta. The subsequent full run records all four Solve completion regressions as recovered in the latest measurement; the underlying structural refusals remain.

Previous complete measurement: selected native Jacobians

target/msl/multibody-selected-jvp-full passes the local MSL quality gate at commit 45cd3b9054be6b57c80fcceba0feedb22818d315, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The complete 566-model run finishes in 122.61 seconds with eleven requested, effective, and pinned simulation workers. 152 models compare and all 152 remain strict-high (26.86% of 566); seventeen existing reviewed exclusions account for the other simulation completions, with zero missing or nonidentifiable traces. All 17,606 initial channels are high and no trajectory channel deviates. Every model’s simulation band and execution outcome is unchanged from electrical-affine-polished-full, including all nine original electrical counterexamples. The only phase-status change is Inverse_sh_TX returning its earlier Flatten refusal instead of a DAE refusal.

MultiBody remains 19/42 high, with nineteen compared, all 8,477 trajectory and initial channels high, and no skipped, missing, excluded, or nonidentifiable MultiBody traces. DAE completion remains 35/42; the remaining 23 examples have the same failure classifications listed below. RollingWheel now reports 9.409 seconds of simulation including initialization, versus 12.601 seconds in the preceding full run. The selected native JVP change retains every prior high model but does not yet unlock another example. The per-model delta and artifact hashes are recorded in rolling-wheel/selected-jvp-full-delta.json. That run precedes the scalar contact-coordinate repair below; this MSL gate does not establish verify quick or verify full success or complete MultiBody coverage.

Coupled scalar contact coordinates

The outstanding core fixture replaces the direct circular constraint with s+w=x, s-w=y, and 2*(s*s+w*w)=radius*radius. Its source and DAE retain all seven equations, but structural matching previously stops at 6/7. The auxiliary recognizer admits aggregate vector maps and dot equations; it cannot materialize these two separately declared scalar coordinates. The direct-coordinate control already passes. This is the known reduced core failure, not evidence that a remaining MultiBody example has closed.

The repair applies the existing SPEC_0007 STRUCT-T03 square linear auxiliary profile to blocks of source-authored scalar residuals. A closed affine proof admits sums, signs, and products with unknown-independent state/invariant coefficients. Connected components must be square; nonlinear and underdetermined components remain refused. Membership, equation identity, and state anchors belong to one shared block. Its coordinates project one aggregate LinearSolve result, with one function and derivative cache per block. The recognizer never consumes tensor scalar views or enumerates a tensor extent. Source residuals remain owners, and their own reconstruction cannot be used to replace them with identities. This preserves MLS §8.3.1 equation semantics and §8.6 initialization obligations through the existing checked DAE construction and SPEC_0032 aggregate execution.

For A*q=b, reconstruction uses the existing aggregate relations A*q'=b'-A'*q and A*q''=b''-A''*q-2*A'*q'; runtime singularity remains a checked solve failure. The original circular-motion regression now passes with both BDF and RK. Thirteen manifold tests pass, including linear and quadratic time-dependent coefficients, state-dependent coefficients, and rejection of inconsistent fixed initial positions. All 153 structural library tests pass, including shared membership and refusal controls; the complete core suite passes 559 tests with zero failures. Affected all-target/all-feature Clippy checks pass after extracting the component membership update from an excessively nested loop.

Review then finds that memoizing affine decisions alone still copies shared coefficient subexpressions into trees. A focused RED doubles source-DAG depth from four to eight and grows the coefficient recipe from 174 to 2,574 nodes. The corrected recipe shares source expressions and coefficient identities; operand traversal visits each shared recipe once. Reconstruction caches each coefficient with its source identity, coordinate/offset role, derivative order, exact call context, reconstruction mode, and provenance. The DAG-growth regression now passes, and all 154 structural library tests pass. Core and canary validation are repeated for this final representation change: all 560 core tests and the affected all-target/all-feature Clippy checks pass.

An added producer test passes: source DAE has zero aggregate linear solves; the prepared DAE has exactly one, retains seven continuous equation owners and both position/velocity manifolds, and both original scalar definitions still reference both auxiliary coordinates. The worker’s current ir-structural-dae.* writer serializes its input DAE before reduction, so those files are not used as evidence of this producer change.

The OMC stage dump explicitly retains der(s), der(w), and their second derivatives. Its original circular fixture produces twelve rows and agrees with the analytic solution in all seven physical channels; maximum absolute error is 6.485e-8. A separate changing-coefficient fixture fails OMC initialization (0 != 1 = $START.vy-vy) and produces no data rows. It has analytic Rumoca regression evidence only; no OMC trace agreement is claimed for that fixture. An initial test written with powers also exposed the existing structural refusal of u^2; the coefficient regression uses multiplication to isolate this repair, and power differentiation remains separate work. A native worker run of each fixture checks all seven channels at eleven samples through 0.1 seconds against the analytic solution; maximum absolute error is 2.731e-10. The original fixture also matches OMC at all nine exactly shared timestamps, with maximum absolute difference 6.485e-8; two Rumoca timestamps have different floating-point encodings from the OMC grid and are covered by the analytic check only.

The fixed target/msl/multibody-scalar-contact-canary passes with all twenty phase, simulation, initial-condition, and band outcomes unchanged from multibody-selected-jvp-canary. Nine models compare high and all 175 initial channels are high, with zero skipped, missing, excluded, nonidentifiable, or deviating traces. At HEAD 705ad773801b86f51e01c1ba63822ea58e92703a, its working-tree digest is b4737e571d9ded6073b571cc86dd997c1acea27a5d08f620cc0585d9589e69eb. The delta is rolling-wheel/scalar-contact-canary-delta.json; fixture traces and equations are under rolling-wheel/scalar-contact-*. The subsequent complete comparison is recorded in the latest measurement. The repeated target/msl/multibody-scalar-contact-shared-canary retains all twenty outcomes and the same nine compared high models, 175 high initial channels, and zero boundary/deviating counts. Its digest at the same HEAD is f1ab545d8ddecd19f40edda2b3a6cd8bd1c1cc179b9d2214d9c42d04d312934f; the delta is rolling-wheel/scalar-contact-shared-canary-delta.json. Native fixture reruns retain all seven channels within 2.731e-10 of the analytic solution. Formatting and git diff --check pass.

Previous complete measurement: electrical parity recovery

target/msl/electrical-affine-polished-full passes the local MSL quality gate against the checked-in fallback baseline. Its complete 566-model comparison finishes in 132.74 seconds at commit 3a7c0b12f46f94220cd56e7b8f64f0337d18a08a, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35, with eleven requested and pinned model workers. 152 models compare and all 152 are strict-high (26.86% of 566). Of 169 simulation completions, seventeen have existing reviewed exclusions; missing and nonidentifiable traces are zero. There are 17,589 high and seventeen minor trajectory channels, zero deviating or severe channels, and all 17,606 initial channels are high. The tracked baseline and exclusion policy are unchanged.

All twelve high-model execution regressions from the first affine sweep are restored. Every previously high model from the RollingWheel and fixed-pre full runs remains high, and all nine original electrical counterexamples now close as strict-high in this complete comparison:

MultiBody returns to 19/42 high, with nineteen compared, all 8,477 trajectory and initial channels high, and zero skipped, missing, excluded, or nonidentifiable traces. DAE completion remains 35/42. RollingWheel completes its four-second horizon with all 184 channels high; its reported 12.601-second simulation runtime includes initialization and remains close to the unchanged twelve-second integration budget. A perf investigation is next to establish where runtime is spent before claiming a performance repair or attributing earlier timeouts to host load.

The remaining 23 MultiBody examples comprise eleven Solve timeouts, three structural refusals, two simulation timeouts, two Flatten timeouts, four DAE refusals, and one Instantiate refusal. The known implicit-contact-circle core regression still fails at structural matching, 6/7. Combined verify quick and verify full are not green, and no PR, release approval, or baseline promotion is made from this gate alone.

One previously completed reviewed exclusion, RectifierBridge2Pulse.ThyristorBridge2Pulse_RLV_Characteristic, now times out in simulation. This execution loss remains visible even though it was never strict-high. GearConstraint reaches its structural refusal within the phase budget instead of timing out; Inverse_sh_TX changes from Flatten to DAE refusal without a corresponding producer edit. Those diagnostic changes are recorded, not claimed as new language capability. Full per-model deltas and artifact hashes are in affine-polished-full-delta.json and full-execution-delta.json under the inverter investigation.

The subsequent RollingWheel perf diagnostic captures 4,001 samples with zero lost samples at 199 Hz and completes under the unchanged solver budget. It reports 11.776 seconds including 0.575 seconds of initialization. In the last 45% of the capture, prepared row evaluation accounts for 22.42% of self samples, tensor concatenation 5.82%, memory comparison 5.65%, matrix multiplication 3.59%, and scalar pure-call payload adaptation 3.35%. Source inspection finds selected projection residual/JVP evaluation uses the interpreter even though the backend interface offers compiled whole-block evaluation. Call metadata is also checked in both the interpreter and native adapter. These are investigation leads, not established repair claims; compiled-program availability and selected-output ownership need verification before changing dispatch. rolling-wheel/affine-polished-performance-receipt.json records the profile, request, result, and hashes. No gate result is replaced by this diagnostic run, and host contention remains an unproven explanation for the earlier timeout.

Selected native Jacobian execution

The bounded availability probe confirms that RollingWheel’s implicit residual, Y-only JVP, and full-seed JVP all compile successfully. The first avoidable work is in RefreshProjectionModel::eval_implicit_jacobian_v_row: it always evaluates the selected prepared program in the interpreter. The native adapter already owns one callable per Jacobian program, but its public execution interface previously exposed only a whole-block sweep.

Two focused regressions reproduce the missing dispatch: four selected requests make zero native calls, and a selected native failure is bypassed while the interpreter returns a value. They pass after adding a selected program-output entry point through the runtime, simulation adapter, and Cranelift product. The coordinate comes from the existing prepared JVP view, including its exact program index and aggregate output offset. This follows SPEC_0007 SOLVE-C13/C14/C15, SPEC_0032 scalar-view ownership, and SPEC_0029’s runtime/backend boundary. It changes execution of the existing AD artifact; it does not change Modelica equations, tensor ownership, or differentiation.

The native entry point checks the program, output extent, and input extents, executes that program into reusable private aggregate storage, and returns only after successful execution. Unrelated programs remain unevaluated. External-table context, sparse whole-block output placement, static-gradient reuse, and both seed spaces retain their existing ownership. An unavailable selected entry point returns None; an execution failure propagates.

The library run passes 440 solver, 72 native-backend, and 127 simulation tests. An added mapping test passes with reordered JVP programs and a multi-output JVP program. The thirteen native policy tests also pass after their counting wrapper delegates the new entry point. Native tests cover changing Y/P/time/seeds, sparse output placement, aggregate output offsets, invalid coordinates and extents, and an unrelated missing-table program that must not execute. All affected all-target/all-feature Clippy checks pass. Core validation has 556 passes and the same one contact-circle structural failure, 6/7. Original-model OMC/perf, canary, and full-cohort validation are recorded below and in the latest measurement. The red failures are recorded in rolling-wheel/selected-jvp-receipt.json.

The originating comparison target/msl/multibody-selected-jvp-origin at HEAD 7e05287df7eb7dede5cb95b9101968588f0b443b, working-tree digest 007525460aca735748b223cd2ad120e16fb7252da6e6d0ed39e68e9343ee3bae, compares one model and keeps all 184 trajectory and initial channels high. Skipped, missing, excluded, nonidentifiable, and deviating counts are zero; worst channel bounded normalized L1 is 1.272e-4. The reported simulation runtime is 8.887 seconds including initialization.

A matching standalone perf capture at 199 Hz, with the same model settings, four-second model horizon, and twelve-second solver budget, completes in 8.723 simulation seconds including 0.455 seconds of initialization. The preceding capture reported 11.776 seconds including 0.575 seconds of initialization: a 25.9% reduction across these two captures. Build time remains about four seconds. The new capture has 3,400 samples and zero lost samples. In its last 45%, prepared row evaluation falls from 22.42% to 11.79% of self samples and native Jacobian row functions appear among the largest entries. Residual evaluation and call-interface work remain visible. This is bounded diagnostic evidence, not a statistical benchmark or a new cohort number; the original-model OMC comparison separately establishes trace agreement. No temporary probes remain.

The fixed target/msl/multibody-selected-jvp-canary passes with all twenty phase, simulation, initialization, and model-band outcomes unchanged from electrical-offset-ports-canary. Nine models compare high; all 175 initial channels are high, with zero skipped, missing, excluded, nonidentifiable, or deviating traces. Its working-tree digest at the same HEAD is 35c3a4205ecd131713a68486f37f4104d563cb5a5dd98500d450e4f73596e34a. The durable delta is rolling-wheel/selected-jvp-canary-delta.json. Formatting passes. The subsequent complete cohort at 45cd3b90 uses eleven simulation workers and retains all preceding high models and electrical counterexamples, as recorded in the latest measurement.

Previous complete measurement: first direct affine solve

target/msl/electrical-affine-full completes the full 566-model comparison in 127.82 seconds at commit 82630a4d2361ef673d921c49ada8519ce51ed08b, working-tree digest 48672dfdec9c3416aac00ca9d56725958329a60236f921974bbf966a09020c35. The quality gate fails: 139 models compare and all 139 are high (24.56% of 566), but twelve previously high models now fail execution. The original center-tap rectifier gains high status. Of 150 simulation completions, eleven have existing reviewed exclusions; missing and nonidentifiable traces are zero. Of 14,471 trajectory channels, 14,461 are high and ten minor; none deviate. All 14,471 initial channels are high. Seven previously observed exclusions no longer complete; the exclusion policy itself is unchanged. These failures are not closures.

MultiBody remains 18/42 high, with 18 compared and zero skipped, missing, or nonidentifiable traces. DAE completion remains 35/42. RollingWheel still times out with eleven pinned workers. Its performance investigation remains pending while the electrical execution regressions take priority.

The numerical follow-ups below restore all twelve lost high models in focused comparisons, but a new complete measurement is still required. Neither this full run nor the follow-up establishes release readiness.

Previous complete measurement: fixed-pre iteration

target/msl/electrical-fixed-pre-full completes the full 566-model MSL/OMC comparison in 149.53 seconds at HEAD 49477020cc7706352fa314fe8384c08d9695b1d8, working-tree digest ee6f7e2ea0ff9fd0f00ee2d581f7340c07b021e55635f049c1dcd961a9130d46. Of 168 simulation completions, 150 compare and all 150 are strict-high (26.50% of 566). There are zero near models and zero deviating or severe channels among 17,115 compared trajectory channels; all 17,115 initial channels are high. Eighteen skipped traces have existing reviewed policy exclusions; missing and nonidentifiable traces are both zero. The exclusion policy is unchanged: RectifierBridge2mPulse.ThyristorBridge2mPulse_RL now completes and enters its already tracked exclusion, increasing the observed excluded count from 17 to 18.

Eight of the preceding nine electrical counterexamples now compare high. The remaining RectifierCenterTapmPulse.ThyristorCenterTapmPulse_R fails with EX002: event condition equations did not converge with fixed pre at t=0.00333333333333337. This is an execution regression, not counterexample closure. It also loses baseline-certified high status, so the quality gate fails. Investigation remains focused on this switching event.

MultiBody measures 18/42 strict-high, with 18 compared and zero skipped, missing, or nonidentifiable MultiBody traces. The previously high elementary RollingWheel now exceeds the unchanged 12-second solver budget. Its simulation timing is 12.598 seconds including initialization; the remaining 24 examples have no comparable successful trace. DAE completion stays 35/42. The scheduler requested 16 workers and memory-capped execution at 11 pinned workers, which the user accepted. The cause of the RollingWheel timeout has not yet been profiled; the preceding 19/42 result is historical. Budgets, thresholds, references, and the baseline were not relaxed.

Fixed-pre event equation investigation

The originating Rectifier1Pulse.Thyristor1Pulse_R_Characteristic has the source equation off = s < 0 or pre(off) and not fire. Its source, DAE, and Solve program retain that equation and the separate pre(off) slot. The old runtime turned the thyristor on at time 0.020000000000019592, before its correct firing edge at 0.02002002002004838. OMC changes both off and fire at 0.02002002002102569. At time 5.005, Rumoca previously conducted while fire was false and OMC held the device off. The final firing pulses agree; the defect is not a shifted controller waveform or lost source operator.

A periodic timer/reset/latch reduction reproduces the defect with both BDF and RK: at time 1.01 the latch wrongly holds true, before the required 1.1 firing edge. A one-shot control already passes. OMC agrees with the periodic fixture’s analytical latch, timer, and integral, with maximum integral error 2.494e-13. Both solver regressions now pass. The governing contracts are MLS Appendix B, SPEC_0022 SIM-001/SIM-008/SIM-009, and SPEC_0040 SOLVE-C22: solve current discrete and condition equations with fixed pre before advancing ordinary event history atomically.

The runtime had refreshed condition memory only after settling discrete equations, then advanced pre before the changed conditions could correct a transient latch value. The new inner iteration settles current equations and condition memory together under the same history snapshot. It retains the clocked first-pass restriction and fails explicitly on nonconvergence. The original focused MSL comparison passes on all 95 trajectory and initial channels, with zero skipped, missing, excluded, or nonidentifiable traces. Its regenerated OMC trace is byte-identical to the preceding reference.

The fixed 20-model canary retains every preceding phase, simulation, initialization, and band outcome: nine compared, nine high, all 175 initial channels high, and zero skipped, missing, excluded, or nonidentifiable traces. Solver tests pass 434/434. Core tests pass 552 with the one known implicit-contact-circle structural failure; focused all-feature Clippy passes. Combined verify quick and verify full are not green. Evidence, artifact hashes, and the full delta are under buffered-relation-counterexamples/thyristor-1pulse in the campaign directory.

Remaining rectifier: stale affine coordinates

The next investigation maps the center-tap failure to thyristor 3: its off value is Solve P230, its switching variable s is Y108, and root 5 writes relation-memory P383. Its 39-variable algebraic block tears on Y113, the common output voltage. OMC preserves the same off equation as regular equation 329 inside coupled block 354. Rumoca’s 32 fixed-pre iterations alternate off between true and false: the projected s alternates between 3.7526e-15 and -9.0816e-14. No root overrides remain during the oscillation.

At the failing time, a 70-digit solution of the original resistor network using its binary64 source voltages gives s = -1.81523e-12 on the off branch and s = -9.08160e-14 on the conducting branch. Both are negative, so the all-off branch is consistent. The runtime instead retains the preceding conducting output voltage after changing branch: its residual and correction are both below tolerance, but the resulting switching variable has the wrong sign. The first divergence is the algebraic projection, not the source equation, event history, or a physically unsolvable switching state.

A two-variable affine regression reproduces the admission error without MSL: a certified solve returns the incoming positive 1e-20 instead of the unique negative -5e-21 solution. A source reduction using three instances of a scalar switch also exposes BDF’s delayed turn-off at time 0.10001431503167271. Both BDF and RK now match the analytical switching state, current, and current integral. OMC matches the same fixture on 204 rows, with maximum current error 1.089e-13 and integral error 1.302e-14. Two initial reduction variants encountered separate DAE array pre shape and projected Boolean-assignment refusals; those are preserved as distinct failures, not evidence of this runtime defect.

Construction-certified affine blocks now compute A*x = -F(0) and validate the resulting residual using the existing tolerance. Exact singleton assignments retain their direct path; nonlinear blocks retain their existing branch-preserving iteration. This uses the existing block-affinity proof, Jacobian, and factorization cache, without changing canonical IR, relation thresholds, event budgets, or comparator policy.

The original center-tap model passes the focused run target/msl/electrical-affine-origin at HEAD 49477020, working-tree digest 913bd843d9ec054f9f969571429aca3317a58a221bf2a7a56608cd8ecec6ad4a: one compared model, all 307 trajectory and initial channels high, zero skipped, missing, excluded, or nonidentifiable traces. Worst channel bounded normalized L1 is 9.174e-6. This is focused evidence; the latest complete cohort above still records its failure. Projection tests pass 60/60, event integration tests 54/54, and core tests 554 with only the known implicit-contact-circle failure. The existing affine refresh test is strengthened from accepting a small-residual iterate to requiring its unique exact zero solution. Evidence is under buffered-relation-counterexamples/center-tap-mpulse.

The fixed 20-model canary target/msl/electrical-affine-canary retains every preceding phase, simulation, initialization, and band outcome: nine compared, nine high, all 175 initial channels high, and zero skipped, missing, excluded, nonidentifiable, or deviating traces. Its working-tree digest is ec3bde07fcd915bdebb58cff20412e7250d4ac6ba2d04cb389cbf6ee809f8240 at the same HEAD. Solver tests pass 435/435; focused all-feature solver and core Clippy pass. The full 566-model sweep is next, with 11 simulation workers explicitly accepted by the user. No full-cohort closure or release readiness is claimed from these focused results.

Affine residual scaling and refinement

The full sweep above exposes two defects in the direct affine path. In ChopperStepDown_R, its first solution has opposing currents near three million amperes during a switching iteration. The candidate’s flow residual is 1.979e-10, with a normalized residual of 1.184e-16 under the existing coordinate-scale policy. The new path had instead certified it using the temporary zero origin’s scales and rejected it against 1e-10. A three-equation resistor/current-source reduction reproduces the rejection. Certification now derives scales from the actual candidate, as the existing Newton path does; no model tolerance or comparator threshold changes.

In RectifierCenterTap2mPulse.ThyristorCenterTap2mPulse_R, the corrected scaling still exposes a first-solve residual of 1.56146e-10 in a small current, just above the unchanged 1e-10 tolerance. A four-variable circuit with large opposing currents and small leakage reproduces that refusal. The affine solve now refines against the original residual using the same certified matrix until the normal certificate passes or the existing projection iteration budget is exhausted. Failure preserves the incoming coordinate. Both numerical regressions turn from red to green; all 437 solver tests and focused all-feature Clippy pass.

target/msl/electrical-affine-refinement-focused compares 14 of 15 selected models, all high, with 3,340 high and three minor trajectory channels, all 3,343 initial channels high, and zero skipped, missing, excluded, nonidentifiable, or deviating traces. Its HEAD is 82630a4d, working-tree digest b7d918c8e691a3b53dd091e6ac9b7fe5c99ec8455a9f8a63baecbfe09e1670b2. All nine original electrical counterexamples are high in this focused check. Eleven of the twelve execution losses are restored; the remaining DCAC.SinglePhaseTwoLevel.SinglePhaseTwoLevel_RL fails condition iteration at time 0.0009362657394780009. It remains an execution regression requiring repair, and the focused gains do not replace the latest full-cohort counts. Artifacts and the two red regressions are under the affine-regressions subdirectory of the center-tap investigation.

The follow-up canary target/msl/electrical-affine-refinement-canary keeps all twenty members’ phase, simulation, initialization, and band outcomes unchanged: nine compared and high, 175 initial channels high, and zero skipped, missing, excluded, nonidentifiable, or deviating traces. At the same HEAD its working-tree digest is 8db4964fc62f6158d73caf1c5118c43a19c4478185979bcc4b6936f96c2adacf. Core tests pass 554 with the same one implicit-contact-circle failure; focused all-feature Clippy and formatting pass after extracting the finite correction update to keep nesting within the existing limit.

Inverter commutation beside offset port voltages

The remaining DCAC.SinglePhaseTwoLevel.SinglePhaseTwoLevel_RL failure is a lower diode-to-transistor conduction transfer, rather than a PWM firing change. OMC’s regular equations 246/249/252/255 retain the same four switch relations. At time 0.0009362657637039185, OMC transfers conduction with lower firing true and upper firing false. Rumoca’s source, DAE, and Solve preserve these equations; the failing runtime state has inductor current 0.0020000011485478463 at time 0.0009362657394780009.

The lower transistor’s switching coordinate is Solve Y43, root 2 writes condition memory P251, and its source off is P168. Fixed-pre iteration alternates the root between zero and -7.1054e-10. A 70-digit solution of the source resistor network, using the captured binary64 inputs, instead requires a negative junction voltage in both candidate configurations: -1.14855e-14 on the transistor-off branch and -5.74274e-15 with both lower devices on. Absolute port potentials near -50 V cause factorization to lose the small independent junction coordinate. Its residual is already within tolerance, so the prior refinement path accepts zero before trying a correction. This localizes the divergence to runtime affine projection; the source relations, history ownership, and upstream IR remain correct.

A four-variable numerical regression reproduces the failure with two offset ports, junction voltage, and leakage current. It formerly returned zero instead of the analytical negative voltage near -1e-17. The general affine path now attempts one residual correction before accepting a nonzero first residual, even when that residual already fits tolerance. Exact zero requires no correction, finite stagnation retains the existing certificate, and nonfinite corrections fail without publishing the private candidate. Subsequent certification and iteration budgets remain unchanged. This implements the existing MLS Appendix B and SPEC_0040 SOLVE-C22 obligation to solve the current equations before accepting their relation values.

All 438 solver tests pass. An additional timed source control verifies inductive current and complementary switching with both BDF and RK; OMC matches its analytical current within 8.676e-11 across 84 rows, with zero switching errors. The complete core suite passes 556 tests with only the known implicit-contact-circle structural failure, 6/7 matched rows. Temporary runtime probes are removed. These checks do not establish verify quick or verify full success.

The originating focused comparison target/msl/electrical-offset-ports-origin at HEAD c4789fce2e90aaa0b33f1406007a6a90642ada43, working-tree digest 2833c01141a5261c5c48c7aa9f5c3301be1e661ab2e1f75e9ea998e54857c991, compares one model and classifies it high: 139 trajectory channels are high and four minor, with all 143 initial channels high. Deviating, skipped, missing, excluded, and nonidentifiable counts are zero. Runtime simulation takes 0.868 seconds. Proof, red/green logs, source/IR hashes, and the OMC control are under buffered-relation-counterexamples/inverter-rl in the private campaign directory.

The fixed target/msl/electrical-offset-ports-canary retains all twenty phase, simulation, initialization, and model-band outcomes from the preceding refinement canary. Nine models compare high, with all 175 initial channels high and zero skipped, missing, excluded, nonidentifiable, or deviating traces. At the same HEAD its working-tree digest is dfec1300d8534c5246c51b960f04ee0b4a3891470b1a1fabdee91d6146cd3570; the delta is offset-ports-canary-delta.json. Solver/core all-target, all-feature Clippy and formatting pass. The complete 566-model sweep is next, using the user-approved eleven simulation workers.

Previous complete measurement: RollingWheel completion

target/msl/multibody-rolling-wheel-full completes the full 566-model MSL/OMC comparison in 298.66 seconds at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, working-tree digest be5115934a4405aed33efc58bb7e96e93cd1f0f0cabd031311a603c7a50d1312. The quality gate still fails on the nine electrical counterexamples below. Of 169 simulation completions, 152 compare: 143 are strict-high (25.27% of 566), and nine are near. Seventeen unchanged reviewed exclusions account for all skipped traces; missing and nonidentifiable traces are both zero. The 59 deviating trajectory channels, including five severe, are unchanged from multibody-torn-roundoff-full. All 17,606 compared initial channels are high. These are comparator bands; a near model can contain deviating channels.

MultiBody now has 19/42 strict-high examples, up from 14/42. All 19 completed traces compare, with all 8,477 trajectory and initial channels high, and zero skipped, missing, excluded, or nonidentifiable MultiBody traces. The five gains are HeatLosses (798 channels), PendulumWithSpringDamper (422), RollingWheel (184), SpringDamperSystem (605), and Fourbar2 (585). The same compiler changes also add high-band results for magnetic MovingCoilActuator.ArmatureStroke, rotational RollingWheel, and translational InitialConditions. ArmatureStroke has two minor channels and zero deviating channels under the existing high-band policy. All 135 models high in the preceding full sweep retain that band. No thresholds, budgets, exclusions, or baseline were changed.

The remaining 23 MultiBody examples comprise eleven Solve-stage timeouts, two simulation timeouts, three structural refusals (PrismaticConstraint, DoublePendulumInitTip, and GearConstraint), two Flatten timeouts, four DAE refusals, and one Instantiate refusal. All 42 remain in scope. DAE completion remains 35/42. Fourbar_analytic advances from structural refusal to a Solve timeout, while RollingWheelSetDriving advances from a Solve timeout to a simulation timeout. PrismaticConstraint now returns its structural refusal within budget. These failure-stage transitions are not parity gains.

All fixed 20 canary members retain their preceding phase, simulation, initialization, and band outcomes when selected from this full sweep; nine compare high, including all 175 initial channels, with zero missing, skipped, excluded, or nonidentifiable traces. This subset comparison is a regression tripwire, not a separate canary execution or cohort measurement. The full-run delta and artifact hashes are recorded in rolling-wheel/full-msl-rolling-wheel-delta.json under the campaign directory.

The nine actionable counterexamples remain in Modelica.Electrical.PowerConverters.Examples.ACDC:

Every row remains non-high, including rows with only minor channel errors. These block merge/release and further MultiBody breadth. Original RollingWheel is now repaired in the full cohort, so electrical triage resumes next. The earlier four electrical closures remain high. Combined verify quick and verify full are not green; the known implicit-contact-circle unit regression also remains unresolved. The requested checkpoint commit preserves these failures explicitly; it is not release approval or baseline promotion. The completed run used four pinned model workers. The user explicitly requests approximately 16 simulation cores next time, overriding the local four-worker preset for that future execution; this run was left unchanged.

Previous complete measurement and RollingWheel investigation

target/msl/multibody-torn-roundoff-full completes the full 566-model comparison in 312.72 seconds but fails the quality gate at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, dirty-tree digest f50ac09f35fdac3c6c65964542478e8ebdb0cf896e54a2176d257d880447ddad. Of 161 simulation completions, 144 compare: 135 are strict-high and nine near. Seventeen reviewed exclusions remain visible; zero traces are missing or nonidentifiable. There are 59 deviating trajectory channels, including five severe. All 14,721 initial-condition channels remain high. MultiBody retains 35 DAE completions and 14 high models out of 42; its remaining outcomes are seven compilation failures, nine solver failures, and twelve timeouts.

The recovered-coordinate roundoff repair closes four counterexamples in this complete cohort: DiodeBridge2Pulse, HalfControlledBridge2Pulse, ThyristorBridge2Pulse_R, and HBridge_R. No previously high model loses its band. DCPM_Cooling remains a timeout. GyroscopicEffects changes its timeout classification from an internal simulation timeout to a parent model-attempt timeout, without a trace. Inverse_sh_TX advances from Flatten refusal to DAE refusal without an owning-phase change in this repair; that classification delta is retained and is not claimed as a capability improvement.

Before the new RollingWheel regression, all 434 solver and 526 core tests passed, as did affected all-feature Clippy checks and formatting. The original DiodeBridge2Pulse focused comparator has all 91 trajectory/initial channels high, one comparison and zero missing, skipped, excluded, or nonidentifiable traces; its OMC reference is byte-identical to the prior full run. The fixed 20-model canary retains all phase/band outcomes, nine compared models and 175 high channels including initial conditions, with zero missing or skipped comparisons. Both use the same dirty digest as the full run. Deltas and root-cause evidence are under buffered-relation-counterexamples/diode-bridge/ in the campaign directory.

Nine electrical counterexamples remain actionable release blockers. The user’s latest priority permits fixing one MultiBody model, then requires returning to those electrical counterexamples before further breadth. The selected model is RollingWheel: its balanced DAE fails structural matching at 907/911. The investigation maps unmatched equations and variables, state selection, initialization, and differentiated constraints against OMC before choosing a repair. The 28 unsupported examples comprise seven structural refusals, twelve Solve-stage timeouts, two simulation timeouts, two Flatten timeouts, four DAE refusals, and one Instantiate refusal. All 42 examples remain visible, together with the broader 111 ModelicaTest and 171 nonpartial library inventories. Current combined verify quick and verify full are not green; no baseline, commit, or PR is made.

The RollingWheel investigation now includes OMC’s flattened equations, backend index-reduction dump, selected states, and generated initialization system. Rumoca’s first unmatched contact row belongs to delta_0 = r_road_0 - frame_a.r_0; the whole-family differentiation proof stops at the implicit road coordinate s. A subsequent capture of the actual matching corrects that initial interpretation: unmatched scalar row 814 is the third component, which reads r_road_0[3]=0, not s or w. Its alternating-path closure contains 26 rows over 25 unknowns: the three contact constraints, implicit delta_0 components, orientation, and normalization. Rows 812 and 813 determine the other road coordinates outside this closure. The proof must respect component demand as well as coupled constraints. OMC differentiates a coupled contact constraint set, including intermediate algebraic coordinates, normalized directions, and second derivatives. The initial state-count difference alone is not proof of a lowering defect.

The new implicit_algebraic_contact_coordinates_preserve_circular_motion regression is intentionally RED: replacing the passing circular constraint with s+w=x, s-w=y, and 2*(s*s+w*w)=radius*radius gives EL005, six of seven equations matched. OMC simulates the same source; all six observed coordinates agree with the analytic circle over 12 rows, with maximum absolute error 3.26e-8. This isolates the inability to handle implicit intermediate coordinates, but does not establish that a repair to this fixture alone closes RollingWheel, or that its implicit road coordinates cause the first unmatched scalar row. Exact source spans and OMC regular equation owners are mapped in rolling-wheel/coupled-constraint-source-map.json under the campaign directory. All temporary probes are removed.

A second reduced source isolates selection through a whole-array definition: road={s,w,x*x}; p=road[3], with nonlinear algebraic equations for s and w, and derivative equations for p and its velocity. It first fails at eight of nine structural matches. Structural admission and reconstruction now use the same source-bound element projection through literal array constructors and checked causal definitions. They preserve the original array equations, do not freeze tunable indices, and do not bypass function calls or their derivative annotations. The selected position, velocity, acceleration, and unrelated algebraic channels now match the analytic solution with both BDF and RK. OMC agrees over 12 rows and nine channels to 2.23e-16 maximum error.

All four new source regressions pass, including matrix selection, inconsistent fixed initial values, and a retained assertion in an unselected element. All 146 structural tests pass. The component-change core run has 530 passes and one failure: the earlier implicit-contact circle is still RED. The fixed 20-model multibody-component-differentiation-canary passes with all phase, simulation, and band outcomes unchanged; nine models compare, with zero missing, skipped, excluded, or nonidentifiable traces and all 175 initial channels high. Its dirty-tree digest is 00a3fa1dd1b29a458dc7349ffffc52653e28e0052dda1203b3a2ac8fc85da186. The rebuilt original RollingWheel worker still reports EL005 at 907/911; its Flat and DAE artifacts are byte-identical to the preceding capture. RollingWheel’s coupled contact constraints remain unresolved, and this component repair does not establish another supported MultiBody model.

A third reduced source keeps a circular-motion constraint inside one mixed vector equation: road={s,w,x*x+y*y}; {s,w,road[3]}={1,2,radius*radius}. It first exposed an earlier Flatten defect: the second vector equation had scalar count one instead of three, yielding ED001 at eight equations and ten unknowns. Equation shape inference now consumes the existing Resolve/Instantiate declaration-based scalar proof, which was already used by dimension evaluation. Unknown references still remain unknown. The minimal array-equation test is RED before this change and GREEN after it; all 631 Flatten tests pass. With only the shape repair, the mixed circle reaches a balanced DAE and fails at structural matching, nine of ten. OMC simulates the identical reduced source over 12 rows; its ten observed channels match the analytic solution to maximum absolute error 1.27e-7, including exact algebraic values for both road coordinates.

The shape-change core run retains 530 passing tests, with only the implicit contact and mixed-vector circle regressions RED (zero ignored or filtered). The fixed 20-model multibody-scalar-shape-canary also passes with all phase, simulation, and band outcomes unchanged, nine comparisons, no missing or skipped traces, and all 175 initial channels high. Its dirty-tree digest is 3aed21fa3f16d017bc663c79c5cbc2301fff033dc7716fc811fb559bb42b9b63. These are focused tripwires; the complete-cohort numbers above remain the latest cohort evidence.

The subsequent tensor constraint repair makes that mixed circle pass with both BDF and RK, including all ten analytic channels. The same equations written as scalar equalities serve as a passing diagnostic control. A source-bound component plan now supplies both the differentiability proof and reconstruction. It retains the tensor owner, domain, body shape, and untouched components through a checked array update; only the selected component changes, and its position and velocity constraints remain on the initial manifold. Reduction identities include the owner, body, and component. This is an implementation within tensor owners, without scalar equation owners or changes to the tensor-native requirement.

The first reconstruction exposed a redundant differentiated causal definition and incorrect array-update incidence. Component admission now rejects identical source projection trees before differentiation. DAE dependency projection reads the replacement only at selected coordinates and reads the old tensor at other coordinates; static whole-axis, scalar, slice, and record field selections have focused regressions. A parameter-index RED test also proves that default parameter values cannot justify removing a possible dependency. Only constant selections receive exact selection; tunable or runtime indices retain their possible dependencies.

The tensor owner, middle-component, consistent and inconsistent initialization, and analytic simulation controls pass. The first complete core run after this repair has 535 passes and only the implicit-contact circle RED, with zero ignored or filtered tests. After the parameter-index guard, all 36 DAE evaluator and 146 structural tests pass; the final core run retains 535 passes and that one known RED. Affected all-target/all-feature library Clippy and all-feature core Clippy pass. The fixed 20-model multibody-tensor-constraint-canary passes with all phase, simulation, and band outcomes unchanged from the shape canary: nine comparisons, no missing, skipped, excluded, or nonidentifiable traces, and all 175 initial channels high. Its dirty-tree digest is f60d484b24a4ac81748f927d9a5d6db54e330d3bf04de2c30e0c8bbfcd765ec6. The rebuilt original RollingWheel still fails EL005 at 907/911, with identical Flat, DAE, and returned structural DAE artifacts. The tensor repair therefore adds no demonstrated MultiBody support yet. The source and producer receipts are in rolling-wheel/tensor-constraint-triage.md and rolling-wheel/tensor-constraint-producer-receipt.json under the campaign directory; temporary probes are removed. No further MultiBody breadth precedes the electrical fixes after the selected RollingWheel repair.

The next regression now retains the actual contact geometry in tests/fixtures/index_reduction/TensorContact.mo: a rotating axis, normalized cross product, implicit road/contact vectors, and the same three scalar contact constraints. The independent angle is prescribed; a vertical reaction replaces the full rigid-body dynamics. OMC runs this exact final source without warnings, selects only the angle as a state, and agrees with all 24 analytic observables over 12 samples to maximum absolute error 7.78e-16. Its backend dump differentiates the coupled auxiliary coordinates to first and second order.

Rumoca constructs 24 equations and 24 unknowns but matches only 23 on both BDF and RK. The unmatched row is delta[3] = road[3] - z; its alternating closure contains 17 equations and 16 unknowns. This corresponds to RollingWheel’s third contact-offset row 814 and its 26/25 closure. The same reduced source with the three contact constraints explicitly solved for s, w, and z passes all 24 analytic channels on both solvers. These three focused tests therefore give two RED implicit cases and one GREEN control, after the preceding 535-pass core checkpoint. They remain unresolved regression obligations.

This isolates a coupled constraint limitation beyond component selection: the current reconstruction requires state-based expressions for auxiliary values and retained first derivatives, whereas the contact block defines those quantities implicitly. The next structural change needs a proof for the coupled block while preserving tensor owners and initialization constraints. General dummy derivatives would also require the scope amendment stated by SPEC_0007; no such cutover has been implemented here. Current-source evidence is bound by rolling-wheel/tensor-contact-receipt.json. Two earlier fixture variants separately exposed Integer-to-Real cross admission and dependent fixed-start preservation limits; both are retained in the private evidence, without weakening either checker. Full verification remains outstanding.

The next building block is now a checked whole-tensor LinearSolve operation in Solve schema 70. Construction requires a positive square Binary64 Real matrix and a matching Real vector, and wire replay derives the result type through the same constructor. Its directional program applies A*dx = db - dA*x with aggregate operands and results. Interpreter and native execution use the existing pivoting kernels and report singular or non-finite systems as errors. A scaling regression exposed absolute pivot cutoffs that rejected a well-conditioned system after multiplying both sides by 1e-20; the kernels now reject zero or non-finite pivots instead. Trace comparison tolerances are unchanged.

The affected suites pass (292 Solve IR, 190 evaluator, 69 native tests, plus doctests), as does affected all-target/all-feature Clippy. The core checkpoint is 536 passes and the same three unresolved implicit-constraint failures, with no ignored or filtered tests. OMC solves the time-varying matrix fixture at scales 1e-20, 1, and 1e20; all six analytic channels over 12 samples agree to maximum absolute error 8.89e-16. The fixed 20-model canary at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, dirty digest b762c8cde08e72fb6bc9b5a2a6ca284daf695a66b793e2a4dda02127d85533b4, retains every preceding phase, simulation status, and band. Its nine compared models are high, with zero missing, skipped, excluded, or nonidentifiable traces. This is Tier 1 evidence only, recorded in rolling-wheel/tensor-linear-solve-canary-delta.json; it adds no cohort claim. The coupled structural proof and RollingWheel repair still remain to be done.

DAE schema 37 now carries the internal auxiliary LinearSolve with a constructor-derived Real vector result. Function lowering produces one checked tensor kernel; the extent-4096 tests retain three registers and four operations. Dependency projection includes both complete operands, and unsupported scalar, compile-time numeric, and GALEC paths refuse the intrinsic explicitly. This does not add a Modelica source builtin or complete coupled index reduction. The final bridge suites pass: 187 DAE unit tests, 17 DAE integration tests, eight compile-fail doctests, 37 DAE evaluator tests, and 120 Solve-phase tests. Affected all-target/all-feature Clippy and formatting checks pass.

The bridge canary at the same HEAD, dirty digest 039d158b9cec066a1bdd2c6f5010f339bb22d9875c5316caad59e3df0ddb847b, retains all 20 phase/status/band results from the preceding kernel canary. Nine models compare high, with all 175 initial channels high and zero missing, skipped, excluded, or nonidentifiable traces. The delta and evidence hashes are in rolling-wheel/tensor-linear-solve-dae-canary-delta.json and rolling-wheel/tensor-linear-solve-receipt.json. The structural proof still needs coupled auxiliary definitions and general higher derivatives of the contact geometry; no original MultiBody model is newly supported by this work.

Coupled tensor auxiliaries and shared equation normalization

The reduced contact now simulates on BDF and RK with all 24 analytic channels correct. A two-dimensional block and a varying-coefficient variant retain all six analytic channels, variable/state counts, and original equation-owner counts. Square source-authored dot blocks supply one aggregate A*q=b proof; first and second derivatives use A*dq=db-dA*q and A*ddq=ddb-ddA*q-2*dA*dq. Product, quotient, trigonometric, square-root, cross, and outer-product rules retain tensor operations. Nonlinear coefficients remain outside this proof, and singular matrices produce a checked error.

The source-value proof now retains exact state dependencies through algebraic coefficient definitions. Losing those dependencies had caused an unnecessary orientation-state lift whose zero initial guess made the contact solve singular. The repaired reduced model selects its original angle/height anchors directly. Original function identities also rebuild through their explicit mapping when an auxiliary function is inserted before them.

RollingWheel’s radius equation is written 0 = radius - dot(...). A regression proved that the initial block recognizer missed this equivalent equation form. STRUCT-T08 now specifies one shared signed-zero equation normalization layer. Causal definitions, derivative definitions, state constraints, algebraic lifts, and auxiliary blocks consume its borrowed equality operands. It preserves the original numerical residual, provenance, tensor domains, and source owners; no division, cancellation, reassociation, or effect-erasing rewrite is admitted.

Validation: all ten contact/auxiliary tests and all 146 structural tests pass; structural all-target/all-feature Clippy passes. The complete core suite reports 545 passes and the same one existing scalar implicit-circle failure (EL005, 6/7). OMC independently matches the six analytic channels of the constant, varying, and signed-equation auxiliary fixtures to within 1.2e-16. The earlier same-source OMC contact receipt remains the 24-channel reference. A separate valid scalar/array .+ fixture is still refused by ToDAe (ED020), although OMC matches its analytic solution; its private reproduction is retained without a coverage claim.

The fixed multibody-auxiliary-normalization-canary passes at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, dirty-tree digest 30ad2e0ecb5f41e3beb4a08dc213f24e7572fe398f61915097e2fad24b06bd98. All 20 phase/status/band rows equal the preceding DAE-bridge canary. Nine models compare high; missing, skipped, excluded, and nonidentifiable counts are zero, and all 175 initial channels are high. This is Tier 1 evidence only. Receipts and hashes are rolling-wheel/auxiliary-normalization-receipt.json and rolling-wheel/auxiliary-normalization-canary-delta.json.

The fresh original RollingWheel worker still fails at 907/911 matching (1.48 seconds compilation, 2.44 seconds Solve preparation). Its three dot rows are now recognized, but reconstruction of their orientation coefficients is not yet proved through function calls. The source DAG reaches the axis-selection conditionals in Frames.TransformationMatrices.axisRotation through Frames.axesRotations; a focused function-based contact reproduction is staged for the next proof. No original MultiBody model is newly supported. Quick/full remain unpassed; the priority remains RollingWheel, then the nine electrical counterexamples, before additional MultiBody breadth.

Exact function selection and source-state anchors

Shared function substitution now selects static branches under the exact caller arguments, including literal array selectors, while leaving runtime coordinates and tunable parameters unknown. A reduced function-based contact fixture passes BDF and RK on all 24 analytic channels. Two further regressions distinguish a causal definition from an independent constraint: substituting a definition into itself must not erase its value equation, and following a definition from an independent constraint must retain the original state dependencies.

The reduced rolling contact with both no-slip equations now passes both solvers on all 36 analytic channels. Same-source OMC checks all 36 channels to within 9.4e-9. All 41 eval-dae and 148 structural tests pass, as does their all-target, all-feature Clippy. The complete core suite has 547 passes and the same existing implicit-circle EL005 failure; quick/full are still unpassed.

The fixed multibody-definition-anchors-canary passes at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, dirty-tree digest 4d57de86845b09ff78fbf1125d6de9904ed5bcddd6bccba35b9091c6112cea40. All 20 phase/status/band rows match the preceding function-selection canary. Nine models compare high, zero are missing/skipped/excluded/nonidentifiable, and all 175 initial channels are high. This is Tier 1 evidence only; receipts are rolling-wheel/definition-state-anchors-receipt.json and the corresponding definition-state-anchors-canary-delta.json in the private campaign directory.

Original RollingWheel still returns EL005 at 907/911 matching. Its diagnostic Solve preparation takes 19.13 seconds, exceeding the normal 10-second budget. The observer now selects the two actual no-slip constraints and reaches one remaining angular-velocity/vertical-acceleration pair. The contact constraint has a proof only through derivative order one: its angle-rate vector cannot yet be reconstructed from the selected angular-velocity and angle states. OMC differentiates the corresponding radius constraint twice. The next focused repair is the tensor linear map inside Frames.axesRotations, after which the original model must be rerun. No original MultiBody parity gain is claimed.

Tensor angular-rate reconstruction and higher contact derivatives

The original RollingWheel now completes structural reduction, Solve lowering, native execution preparation, and initialization. Its remaining failure is the unchanged 12-second simulation budget; there is no completed trace or new strict-high claim for this model.

Two reduced source models isolate the structural defects. AngularRateContact requires reconstructing a vector of angle derivatives from a state-dependent linear map. Structural analysis now proves that map directly through source array operations and exact function substitutions, retaining aggregate matrix and outer products. It neither enumerates a tensor basis nor replaces the original equation owners. Nonlinear unknown-dependent coefficients are refused. TensorContactAnnotated exposes a separate order-dispatch defect: a supplied first derivative incorrectly prevented a provable second derivative of the function body. Both fixtures pass BDF and RK checks against their analytic channels; OMC agrees with those channels to maximum absolute errors of 4.59e-8 and 5.56e-16, respectively.

The newly differentiated original DAE exposed repeated traversal of shared expression dependencies. Query-local visitation now keys on expression, scalar, record field, and lexical domain point; function argument contexts keep their separate handling. A twelve-layer shared graph previously emitted one dependency 4,096 times. The original structural preparation fell from 54.510 seconds to 0.617 seconds, producing byte-identical DAE JSON with 12,134 expressions, 368 variables, and 255 continuous equation owners.

Validation: 43 DAE-evaluator and 150 permanent structural tests pass, as do all eight contact tests and Clippy for the changed crates. The core suite has 549 passes and the previously recorded implicit-circle EL005 failure. The fixed 20-model canary at target/msl/multibody-angular-rate-canary has no changes in phase, simulation, initialization, or comparison band relative to multibody-definition-anchors-canary. Its nine compared traces remain high; all 175 compared initial channels are high, with zero skipped, missing, excluded, or nonidentifiable traces. The run binds HEAD bc71577f and working tree digest c47d4fdc15130307b1e62c5248438311ff425890b8bf7b5a84e7ce7efe579388. This is Tier 1 regression evidence; the preceding full cohort remains the source of cohort counts. Quick and full verification are not yet green.

Shared differentiation graph and original RollingWheel completion

perf identifies repeated evaluation of expanded contact derivatives as the remaining runtime cost. Plain reconstruction already preserved shared nodes, but differentiation, exact-value substitution, and function instantiation re-emitted shared subexpressions. The scalar/vector scaling regression grows from 61 to 789 resulting expressions when shared source depth doubles from four to eight before the repair. Reconstruction now caches successful results by source expression, complete call context, derivative order, mode, and provenance. It preserves expression sharing without changing equation owners, tensor structure, floating-point operations, or function-call identity. The scaling test now passes, and ScopedContact checks distinct arguments through nested calls against four analytic channels on both BDF and RK.

The unchanged 12-second solver budget now admits the original four-second RollingWheel simulation. The focused comparator and subsequent full cohort both give 184/184 high trajectory and initial channels, with zero skipped, missing, excluded, or nonidentifiable traces. Maximum channel bounded normalized L1 error is 1.1923173202530037e-4. Rumoca and OMC use different state coordinates; trace parity does not establish identical state selection.

The first focused OMC launch separately exposed an empty endpoint-file race: the file existed before OMC wrote its endpoint. The launcher now waits for nonempty publication within the existing deadline and cleans up its own child on startup failure. The empty-file regression fails before the repair and passes afterward. The failed run is retained; the canonical reference stage subsequently compares the unchanged successful Rumoca trace, without rerunning that simulation. The complete cohort above exercises this fix too.

Focused validation has 151 structural, nine contact, and 169 MSL-tooling tests passing. Temporary performance probes are removed. The latest complete core checkpoint predates graph sharing and has 549 passes plus the known implicit-contact-circle failure. The test-helper nesting lint discovered in the graph-sharing Clippy run is fixed. The final checkpoint reruns all 151 structural tests successfully and passes all-target/all-feature Clippy for rumoca-phase-structural and rumoca-test-msl, formatting, and diff checks. No current quick/full pass is claimed.

Preceding parameter-relation sweep and torn-roundoff repair

target/msl/multibody-parameter-relations-full completes the full 566-model comparison but fails the quality gate at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, dirty-tree digest cf88904a126e6030665feb95ab352e0462bfe150eeec74fcc42305c5fa23e1c8. Of 161 simulation completions, 144 compare: 131 are strict-high and 13 near. Seventeen reviewed exclusions remain visible; zero traces are missing or nonidentifiable. There are 63 deviating trajectory channels, including nine severe channels. All 14,721 compared initial-condition channels are high. ForceAndTorque is restored to all 527 trajectory/initial channels high, and MultiBody again has 14 high models out of 42. DAE/balance totals remain 288/276. DCPM_Cooling loses completion to a timeout, so the strict-high total does not increase. PrismaticConstraint changes from structural refusal to a timeout; neither outcome establishes support.

The thirteen remaining counterexamples are power-converter models. The next is DiodeBridge2Pulse: 90/91 channels are high, but ground.p.i has mean absolute error about 5.4e-4 A against OMC’s zero-current solution. Its torn-Newton certification defect is now reproduced and repaired in focused tests; the original complete trace and cohort confirmation remain pending. The complete per-model delta is retained in buffered-relation-counterexamples/force-and-torque/parameter-relations-full-delta.json under the campaign evidence directory. Breadth remains paused; no baseline, timeout, or tolerance is changed, and no commit or PR is made.

At the first bad sample, 0.0102 seconds, the positive DC current-balance residual equals the erroneous ground current, -1.429416585e-4 A. Replaying the same Solve artifact proves that certified observation refresh accepted this point. Its Newton correction, 7.147075780e-16, rounds away in a tear voltage near -9.77; comparing recovered coordinates after that rounded update falsely reports zero change. Reduced-Jacobian sweeps now retain recovered-coordinate sensitivities, so certification also bounds the linearized correction before addition. A stalled line search declines and restores the incoming seed for the existing full-system solve. The exact captured point then yields zero ground current. Temporary diagnostic source probes are removed.

Two reduced regressions first fail on currents of 0.050004445 A instead of 0.05 A; all three new controls now pass, including equation rescaling and continued reduced solving of a well-conditioned loop. OMC’s equivalent source preserves total current exactly across 12 rows, though individual currents differ from the analytic split by 4.44e-6 A in opposite directions; that numerical difference remains recorded. Evidence is in buffered-relation-counterexamples/diode-bridge/torn-roundoff-triage.md under the campaign directory. This is a general numerical projection repair, with no change to equation lowering or trace policy.

Preceding buffered-relation sweep and parameter-rotation repair

target/msl/multibody-buffered-relations-full completes the 566-model comparison but fails the quality gate at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, dirty-tree digest 3ce52135f14b3144926bbe9252d3e80586b0ac43aba597e7385429a239c9b5f3. Of 162 simulation completions, 145 compare: 131 are strict-high, 13 near, and one deviating. Seventeen reviewed exclusions remain visible, with zero missing or nonidentifiable traces. There are 203 deviating trajectory channels, including 17 severe channels. Initial conditions are high for 144 models; ForceAndTorque has 27 deviating initial channels. MultiBody has 35 DAE completions, 14 simulation completions, and only 13 high models out of 42. Its other outcomes are seven compilation failures, eleven solver failures, and ten simulation timeouts.

IdealTriacCircuit is repaired in this full comparison: all 55 trajectory and initial channels are high. Its focused semantic regressions, 1,070 core/Solve/ solver tests, affected Clippy checks, and unchanged fixed 20-model canary pass. However, eight previously high models lose that band: ForceAndTorque and seven power-converter examples. Fourteen actionable counterexamples remain, including the five preceding electrical counterexamples and a newly completed near trace. The next triage is ForceAndTorque’s incorrect initial force/torque resolution; unrelated breadth remains paused. The complete per-model delta is retained in tensor-affine-counterexamples/ideal-triac/buffer-full-delta.json under the campaign evidence directory. Current verify quick and verify full remain outstanding; this run is not release evidence, and no baseline is promoted.

ForceAndTorque’s parameter-rotation regression now has a general DAE fix under validation. The flattened model is byte-identical before/after. Six relations over parameters incorrectly received crossing roots, including the enum tests selecting fixedRotation.R_rel. Solve’s initial parameter seed was correct, but its initialization update read zero-initialized relation buffers and replaced the rotation with the fallback identity. The source event planner was conservative about qualified enum literals; both expression-event and activation lowering now consume the constructed expression variability and preserve parameter-only tests as literal conditions without roots (MLS §3.8.3/§8.5).

Two reduced tests first failed: two parameter roots remained, and horizontal force was zero instead of -500. OMC’s flattened equations and all four sampled rows establish the expected force. The strengthened suite checks all three record-binding branches on Auto/Interpreter and a parameter when activation. All 524 core and 820 DAE/Solve/solver library tests pass, as does affected-library Clippy. target/msl/multibody-parameter-relations-origin restores all 527 original ForceAndTorque trajectory and initial channels to high, with one compared model and zero skipped, missing, excluded, or nonidentifiable traces. It uses the same HEAD and dirty digest aca1281986332acfef3e30a9139e3a223e457992873655b4eb4cd4d6a03b7b27. The fixed multibody-parameter-relations-canary has all 20 phase/band rows unchanged from multibody-buffered-relations-canary: nine compared models and 175 channels high, with zero missing or skipped comparisons. Evidence and the producer/canary deltas are in .git/multibody-campaign/buffered-relation-counterexamples/force-and-torque/. This is focused closure evidence; the next complete cohort comparison must confirm it and measure the remaining electrical counterexamples.

The three targeted FMI assertion tests also pass with package execution enabled (parameter-relations-fmi-2.log, 12.71 seconds): FMI 2/3 schema and VDM checks, FMPy parameter validation and analytic free fall, plus continuously varying predicate and state-event rejection controls. The first general-shell attempt skipped package execution because CMake/FMPy/Java were unavailable; the passing attempt uses the FMI shell and the existing pinned FMPy environment. OMC also accepts all three reduced rotation branches, with maximum force error below 1.2e-13 against their analytic values (omc/comparison.json).

Preceding tensor-affine failure and relation-buffer repair

target/msl/multibody-tensor-affine-full completes the 566-model comparison but fails the quality gate at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, dirty-tree digest 5ab62ca833eef338e1ad7948e0c8d90fdea446ab04f54d7c8093c476d72ca8fb. Of 157 simulation completions, 143 compare: 137 are strict-high and six have ten deviating channels. Four of those channels are severe. Fourteen existing reviewed exclusions remain visible; no traces are missing or nonidentifiable. All initial-condition comparisons are high. MultiBody retains 14 high models out of 42 and all 5,883 shared channels high, without any new completion.

The six actionable counterexamples are IdealTriacCircuit, DiodeBridge2Pulse, HalfControlledBridge2Pulse, ThyristorBridge2Pulse_R, ThyristorBridge2mPulse_RLV, and HBridge_R. ThyristorCenterTapmPulse_R also loses its preceding high result to an event-iteration refusal. Further MultiBody capability work is paused while the first counterexample is traced against OMC. The 300.23-second test and complete per-model delta are retained as tensor-affine-full-1.log and tensor-affine-full-delta.json under the campaign evidence directory. No baseline, tolerance, or exclusion was changed. Current combined verify quick and verify full are not green; no commit or PR is made.

IdealTriacCircuit now has a reduced source-level regression independent of MSL. Both Auto and Interpreter reproduce a negative forwardSwitch while forwardOff remains false at 0.238 seconds. OMC simulates the same source and satisfies both turn-off implications on all 320 rows. The original model’s first lasting divergence is the turn-off near 0.48522 seconds: the preceding implementation and OMC turn off, while the new exact solve reaches zero and retains the on state.

Temporary instrumentation confirms that the component detects the reduced model’s crossing at 0.2363029383566202 with indicator -1.5349943538467414e-7. The source DAE retains both exact relation identities, but Solve assigns no relation-memory target to either. Consequently, there is no relation-memory slot on which event iteration can retain the detected post side. A third RED test records this construction gap. This rejects a missed-crossing explanation and a defect confined to native code generation; the next task is a general buffered-relation lowering repair under MLS §8.5/Appendix B and SPEC_0007/0038/0044. Relation truth must remain distinct from condition-edge history and from the whole Boolean assignment. The probes are removed, and all three focused tests initially failed, providing the pre-repair evidence. Artifacts are under .git/multibody-campaign/tensor-affine-counterexamples/ideal-triac/.

The in-progress repair now allocates relation truth by exact source expression, separately from Boolean results and condition-edge history. Root programs still evaluate the source operands. A shared-expression noEvent regression first failed and now has a distinct literal-evaluation context, including cache scope. The reduced circuit’s persistent latch error is repaired, but its first trace still had two wrong event-left rows. An independent affine-loop test then proved that derivative-coordinate reuse returned a root of zero instead of -5e-6: a small tear residual did not bound the recovered variable’s error. Derivative and root refresh now require existing coordinate-convergence checks. The four switching/root controls pass without tolerance changes.

A fifth RED control showed that non-strict relations were false at exact-zero initialization. Relation-buffer updates now use the same compiler-owned zero orientation as event indicators; all five controls pass. The complete core check found three homotopy branch-selection regressions (517/520 passed). Continuation now settles relation truth and ordinary Boolean definitions at each lambda, reprojects initialization when they change, and restores both numeric and discrete coordinates after a rejected step. All 12 homotopy controls pass. The later full comparison above closes IdealTriacCircuit but exposes further regressions, so MultiBody breadth remains paused.

Previous passing complete measurement

target/msl/multibody-additive-assignment-full passes the complete 566-model MSL/OMC gate at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, dirty-tree digest 8df44f7325df0da96fdf02f5061d3792822f91a563d7cbeace922a1791b2cdad. The 275.22-second test passes the unchanged quality and performance floors. Flatten passes 491 models, DAE construction 288, and balance 276; initialization succeeds for 176 and simulation for 160. The comparator measures 144 models, all high, with 16 existing tracked exclusions and zero missing, nonidentifiable, or deviating comparisons. All 143 preceding high models remain high. DemoPowerSupplyWithBuffer newly passes all 42 compared channels; its preceding failure was an event-boundary algebraic projection refusal at dcPowerSupply.i. Of 14,721 trajectory channels, 14,706 are high and 15 minor; none deviate. All 14,721 initialization channels are high.

All 5,883 compared MultiBody channels are high, with zero minor, deviating, missing, skipped, excluded, or nonidentifiable comparisons. The 28 remaining failures are one Instantiate refusal, two Flatten timeouts, four DAE-construction refusals, eight structural-analysis failures, ten Solve timeouts, and three simulation timeouts. PointGravityWithPointMasses2 refuses its heterogeneous fixed array. PrismaticConstraint reports its structural refusal instead of the preceding Solve timeout; GyroscopicEffects completes Solve and initialization but exceeds the 12-second simulation budget. No new MultiBody completion or coverage of the wider 111-model ModelicaTest / 171-model library scope is claimed.

The current performance gate passes with system-speedup median 1.716 and wall-speedup median 44.839, against baselines 1.559 and 52.45468. It uses 14 single-threaded OMC workers and the unchanged 35% tolerance. The preceding multibody-verify-quick-checkpoint-6 performance failure remains retained; these later standalone MSL passes do not rewrite it. The complete current delta is additive-assignment-full-delta.json under .git/multibody-campaign/gyroscopic-sim-timeout/. Apart from the new high model, all model agreement bands remain unchanged. Inverse_sh_TX now reaches its DAE refusal instead of failing in Flatten; ThyristorBridge2Pulse_DC_Drive reports a runtime refusal instead of its previous simulation timeout. No baseline was promoted, and neither unresolved failure earns coverage credit.

The last combined verify quick run, before the zero-assignment repair, completed in 1,688.257 seconds. Lint, all 28 pinned corpus cases, 243 architecture tests, 17 repository gates, and all 7,395 workspace tests pass, with no workspace tests skipped. Documentation tests pass 30 examples, with 23 existing ignored examples. Its only failed step is the MSL performance gate. The earlier formal-statement registry failures are repaired and pass that combined run. The latest standalone full MSL gate passes, as do 1,409 focused/core/solver tests and affected Clippy; combined verify quick and verify full still require current runs. No commit, baseline promotion, or PR is made. Exact timings and the complete measurement receipt are under .git/multibody-campaign/gyroscopic-sim-timeout/ as verify-quick-6-quick.json, verify-quick-6-receipt.json, and quick-6-msl-delta.json; the delta’s original target/msl/results directory was renamed to the checkpoint above after verification finished.

In progress: affine tensor moment equations

The next GyroscopicEffects equation is Body.mo:261, frame_a.t = I*z_a + cross(w_a,I*w_a) + cross(r_CM,frame_a.f). Canonical program 426, logical row 1399, output 2 lacks an isolator for solver Y871, bodyCylinder1.body.frame_a.f[2]. The exact program and OMC regular equation 1688 differ by zero after the XML torque alias and the two declared zero r_CM components are applied symbolically. Its force coefficient is -r_CM[1]; the instance value 0.125 is not a compiler assumption. This follows MLS 3.6 §§10.3.5, 10.6.4 and Appendix B, with SPEC_0032 compact tensor ownership and SPEC_0036/0043 source-bound assignment proofs.

A minimal cross-product regression initially returns no assignment where force[2] = (torque[3] + r[2]*force[1])/r[1] gives 15, independently of starts at ±1e30. The new TensorAffine certificate retains one checked rule per source producer, compact independent register ranges, and the selected output. Final execution adapters materialize offset/coefficient pairs using the original tensor operations; shared DAG nodes are not unrolled, calls are not replayed with a substituted target, and the original evaluation prefix remains. Scalar products/division, matrix products, cross products, fills, transposes and linear arithmetic share this construction. Dependent products and denominators remain unproved. Dynamic zero/nonfinite coefficients retain implicit fallback; the complete-seed and nonlinear branch guards are unchanged. Solve schema is 69.

The initial library pass covers 475 tests. A source-level TensorAffineMoment regression proves the force seed, rejects forged source-selection/output/cut metadata, and checks seven analytic channels under BDF/RK and Auto/Interpreter. A second RED test exposed materialized coefficient overflow returning -0 while the per-row evaluator declined it. The existing nonfinite-coefficient guard now belongs to the shared materializer, removing the evaluator’s duplicate emitter. The library suite, all 516 core tests, and 429 solver controls pass after that change: 1,420 tests total. Affected all-target/all-feature Clippy passes. OMC regular equation 13 isolates the same force expression, and the normal trace comparator measures all seven trajectory and initialization channels high (score 1.6705e-7), with no minor or deviating channels. The largest absolute analytic error is 1.4753e-6 for OMC’s torque channel with scale 54, and 1.0202e-8 for Rumoca; no solver or comparator tolerance was changed. The fixed 20-model canary retains every preceding phase/status/band row, with all 175 trajectory and initial channels high across nine comparisons and no missing, skipped, excluded, or deviating comparisons. The normal GyroscopicEffects run still times out in simulation. Its DAE is byte-identical to the previous DAE; ten additional exact seeds reduce missing assignments from 26 to 16. A separate 60-second diagnostic completes in 21.81 seconds and compares all 967 shared OMC trajectory and initial channels high, but does not earn normal-budget coverage. The full cohort uncovers the counterexamples reported above, so the capability is not ready to land. Detailed logs and the changing receipt are gyroscopic-sim-timeout/tensor-affine-* under the campaign evidence directory.

Replaceable record dimensions

The PositionControlledDCPM reduction now typechecks for both a default replaceable record and an explicitly redeclared scalar record, without needing the numerical value of asin/acos bindings. The eleven matrix-dimension controls pass. The complete MSL measurement above includes this repair and confirms the recovered Typecheck stage without additional high-parity models.

Source-to-instance inspection also found that a direct component modifier (Holder h(redeclare DriveData driveData[2])) lost the replacement type and shape, while the equivalent extends modifier retained them. Instantiate now applies a declaration-ID-keyed replacement in the current occurrence before array expansion, using the existing cached class template. A reduced sibling control proves that an unmodified instance keeps its original type and rank. Dimension expressions are resolved and evaluated in the enclosing source scope.

A second RED test showed that the first array element lost an each field modifier: the parser had keyed the modifier by driveData[2], mistaking the new declaration’s dimensions for an element selection. The modifier now targets the whole component. OpenModelica confirms the five reduced cases’ record counts and field values, including shared {3,3} and distributed {3,4} values. Another RED parser test proves retention of nested each, final, and redeclare flags. These follow MLS §§7.2–7.3 and §10.1; no model-specific path or changed tolerance is involved.

The shared dimension proof may use a scalar fact only when every observed occurrence of that declaration is scalar. Any array occurrence vetoes that fact, including a zero-length array recorded before expansion removes all its elements. Unapplied nested redeclarations remain explicitly unknown and invalidate descendant shape facts. This does not claim complete support for nested component redeclarations or other unreviewed redeclaration prefixes.

Evidence is retained under .git/multibody-campaign/: the original RED tests, replaceable-record-overlay-probe-2.log/-4.log, component-redeclare-dimensions-modifiers-red-2.log, component-redeclare-dimensions-omc/comparison.json, replaceable-record-consensus-unit-1.log, and redeclare-nested-prefixes-red-1.log/-green-1.log. Final checks pass all 501 core tests and 824 IR/typecheck/Flatten tests, all-target/all-feature Clippy for seven affected crates, and workspace formatting. The broader affected-crate run also passed 1,377 tests with three existing ignored documentation examples before the final marker cleanup.

multibody-replaceable-record-origin restores PositionControlledDCPM from Typecheck ET004 to its earlier Flatten EF024 (motorData is missing structured identity). The other eight original models retain their DAE-construction frontiers. All nine pass Typecheck; none simulate, so this run reports parity unmeasured and exits 1. Its delta is replaceable-record-origin-delta.json.

The fixed canary target/msl/multibody-replaceable-record-canary exits 0 at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, dirty-tree digest e79e68e65df477423e50f5e9ed3a9955605da310cf93108ceb7bc37238884fd0. All 20 model phases/statuses and band rows match the preceding canary. Nine models are compared; all 175 trajectory and initial channels are high, with zero minor, deviating, missing, skipped, excluded, or nonidentifiable results. The durable Tier 1 delta is replaceable-record-canary-delta.json. The complete checkpoint and its outstanding verification failures are recorded above and in verify-quick-checkpoint-5-status.json.

Generalization audit: explicit constraining interfaces

The current audit compares language rules, source identities, and independent record and differential-equation examples. A textual scan of added compiler, evaluator, solver, and simulation lines against the branch base finds no added model-name special cases; that scan alone does not prove semantic generality.

The audit reproduced a defect in inheritance::validate_redeclaration: it read the default component type identity before the explicit constrainedby identity. MLS §7.3.2 requires comparison with the constraining interface. New reduced tests check a record and a model replacement that omit default-only fields, plus negative controls for missing interface members and an omitted explicit constraint. Both positive tests failed with EI027 against the default type, while their resolved constraint identities were correct. The shared validator now reads the explicit constraint identity and uses the default type only when the clause is absent. Missing identities are rejected without name recovery. All four tests pass, including both solvers’ traces, as do all 505 compiler-core tests and 216 instantiation unit tests. Architecture, specification gates, and affected all-target/all-feature Clippy also pass. The fixed canary target/msl/multibody-constraint-interface-canary passes at dirty-tree digest 3a95cca45f5dee3d3d74f846c99d43d19e9959460b0ca55ec602a03f0e5c225b: all 20 phase/status/band rows are unchanged; nine models compare high with all 175 trajectory and initialization channels high and zero missing, skipped, excluded, nonidentifiable, or deviating results. Its durable delta is constraint-interface-canary-delta.json. The complete measurements include this fix and retain all prior high-parity models; the canary does not establish a new cohort parity count.

OpenModelica accepts both positive sources and simulates y = 2*time exactly. Its default frontend also accepts both negative probes, so its acceptance is not used to relax the MLS requirement. The initial oracle runner exits 1 because its expected negative rejections did not occur; both positive simulations succeeded. Evidence is in constraint-interface-omc/comparison.json and constraint-interface-receipt.json. Broader structural-interface limitations remain separate open proof obligations, and no additional MSL support is claimed.

Current semantic repair: computed fixed attributes

The fresh GyroscopicEffects comparison uses the same pinned MSL 4.1.0 sources and OMC a96aa1a-cmake: frontend flattening, backend XML snapshots, initialization dumps, and a successful five-second reference simulation. Both frontends report 2,276 scalar equations and unknowns. OMC’s 16 selected states versus Rumoca’s 18 coordinates with two quaternion norm constraints is not, by itself, a semantic discrepancy. Fixed-rotation inverse matrices also have parameter bindings in Rumoca corresponding to OMC’s initial equations. The DAE Modelica renderer refuses promoted_cat1 in axisRotation; the DAE JSON remains the authoritative Rumoca artifact for this comparison.

The first proven divergence is earlier, in instantiation. Spherical.mo:97 declares w_rel[3] with fixed=fill(w_rel_a_fixed,3). GyroscopicEffects sets w_rel_a_fixed=true for both spherical joints. OMC retains fixed=true on all six scalar components, but Rumoca’s Flat and DAE attributes were absent. extract_attributes only accepted Boolean literals, discarding computed values or allowing an unresolved outer modifier to fall through to a local literal. Under MLS §8.6 this loses required initialization equations; MLS §7.2.4 also requires the modifier’s written scope and overriding precedence.

The reduced regression first fails with enabled.x.fixed=None instead of Some(true). The replacement evaluates uniform Boolean expressions using the instantiation environment, preserving instance-specific parameter values and outer modifier scope. A supplied but undecidable or nonuniform attribute returns EI035 instead of disappearing. Heterogeneous per-component fixed values still need a checked owner; this repair does not claim that capability. PointGravityWithPointMasses2 uses mixed fixed arrays and now refuses them explicitly. ControlledMixingUnit has the same limitation, and Transformer3PhaseYyWithHysteresis refuses indexed HFixed[1]. All three were already failing later stages; this repair exposes the earlier semantic gaps without claiming support for their attributes.

The dynamics regression uses der(x)=-x and y=2*x, with x.start=0 and y(start={1,2,3}, fixed=fill(pinned,3)), pinned=true. Both Rumoca solvers and OMC initialize x={0.5,1,1.5}, y={1,2,3} and follow the analytical exponential solution on all six channels. The nine modifier integration tests and 956 affected library tests pass; six tensor initialization tests pass with the new dynamic case. The real-model DAE changes only ten fixed attributes: the two three-component angular velocity arrays become true and eight arrays become false. All other canonical storage is exactly equal. Solve restores six initial rows (8 to 14); continuous programs and both layouts remain exactly equal. Pure-call site owner ordinals in discrete/event programs shift by two without other program changes. Focused all-target/all-feature Clippy passes. The fixed canary retains all 20 phase/status and band rows: nine models compare high across all 175 trajectory and initialization channels, with no missing or deviating channels. The standard full-cohort gate passes with all 143 high-parity models retained, 16 existing exclusions, and zero missing, nonidentifiable, or deviating comparisons. GyroscopicEffects still times out in simulation. No additional model support is claimed. Diagnostic artifacts are in .git/multibody-campaign/gyroscopic-omc-stages/ and .git/multibody-campaign/gyroscopic-fixed-attribute/.

Current performance repair: shared pure-call interfaces

After restoring GyroscopicEffects’ computed fixed attributes, a 0.01-second diagnostic agrees with a fresh OMC simulation on all 967 shared channels; initial values differ by at most 1.43e-14 on the diagnostic relative scale. A separate full five-second diagnostic completes in 38.68 seconds of runtime under an explicit 60-second budget. The repository comparator reports a bounded normalized L1 score of 1.037e-6 and a worst channel score of 1.841e-5. This is diagnostic evidence only: the standard 12-second simulation budget still fails, and no additional model is counted as supported.

The actual standard worker’s simulation profile identifies repeated typed-call interface equality and payload work. The Solve artifact contains 2,070 primal and embedded directional site records referring to 253 owners. Each issued site deep-copied its immutable type, dependency, projection, and affinity inventories. A reduced storage-sharing test fails before the repair.

Call sites now share those private immutable inventories with their owner. Equality remains value-based; deserialized sites with separate allocations are still accepted only when their complete metadata matches. No runtime value cache, semantic identity, equation, or numerical policy changes. The 33 focused call tests and 522 IR/evaluator/Cranelift library tests pass, including altered dependency/projection/affinity rejection controls and native execution tests. Affected all-target/all-feature Clippy passes. The first normal origin run still fails, reaching a Solve timeout rather than simulation; that failure is retained. In the separately bounded diagnostic, the complete Solve problem is exactly equal and all 2,276 trace channels at 501 samples are exactly equal before and after sharing. Runtime falls from 38.68 to 33.95 seconds in these diagnostic runs; this is not a standard-budget recovery or a cohort performance claim. An explicit BDF short-horizon probe also succeeds, so the initial BDF capability boundary is not the cause of the timeout.

The fixed canary target/msl/multibody-shared-call-interface-canary passes at dirty-tree digest 9994da98fd6de62d7850bb34f93f9b7603d88b13a1400ee5e84ac0ae4ad21548. All 20 phase/status and band rows match the previous canary: nine models compare high, with all 175 trajectory and initialization channels high and zero missing, skipped, nonidentifiable, or deviating comparisons. Its durable delta is gyroscopic-sim-timeout/shared-interface-canary-delta.json. The complete verify quick outcome is recorded above; verify full remains outstanding.

A controlled diagnostic rebuilds the pre-sharing and shared implementations with the same Cargo command, then alternates four executions of each version on CPU 6 with one worker thread. All 24 runs complete, and each model’s entire trace is exactly equal across both versions and all repetitions. Median runtime changes are 84.663 to 80.835 ms for CompareSincExpSine, 731.556 to 723.295 ms for DemonstrateSignalExtrema, and 1.07647 to 1.05386 seconds for DoublePendulum. These three controls do not reproduce the broad slowdown seen in the cohort. They do not discharge the failed performance gate or establish its host-level cause. Both diagnostic binaries, source snapshots, and every attempt are kept in controlled-interface-comparison/; the current source and canonical build artifacts are restored.

A read-only audit identifies the next GyroscopicEffects runtime lead. All 51 algebraic projection stages and both derivative projection stages lack direct seeds for some coupled coordinates, making the runtime’s staged-execution predicate false. It therefore uses the causal sweep followed by full projection plans containing 2,004 and 479 blocks respectively. OMC’s saved compiler dump classifies the corresponding physical rotational systems as linear after reduction, with 39 and 36 unknowns before tearing; mapping those equations to Rumoca’s 79- and 72-row blocks remains a proof obligation. Replaying the saved SolveModel through the phase-owned checked constructor confirms that every one of these 51 algebraic and two derivative projection stages already carries an affine proof. This does not establish nonsingularity: ordinary projection can accept an exactly zero residual before factoring its Jacobian. A staged execution repair must preserve singular-system behavior, nonlinear branch selection, and coordinate certification. No seed check has been relaxed and no speedup is claimed for this hypothesis. Evidence is .git/multibody-campaign/gyroscopic-sim-timeout/, including value-stage-coverage-audit.json, constructed-refresh-facts.json, staged-refresh-candidate.json, and omc-coupled-runtime-blocks.json.

Exact zero-flow assignments

The next producer-level reduction identifies a simpler missing seed before the coupled blocks. GyroscopicEffects logical row 2159 owns bodyCylinder2.frame_b.f[1] (Y1264); its compact program loads three Y values and stores them as residuals. The previous isolator issued AffineResidual, which the declared-target seed constructor excludes. OMC’s saved equations 1323–1325 assign these three forces to 0.0.

Solve assignment construction now issues Zero for an exact target load or certified total copy, including a scalar negation. It retains the original evaluation prefix and target identity. Runtime evaluation and final scalar materialization consume that certificate directly. Dynamic or zero coefficients, nonlinear products, overwritten registers, and unproved calls do not acquire this certificate. Solve schema 67 and checked wire replay bind the new representation. A second RED regression fixes scalar materialization’s register allocator to advance past every tensor lane before appending a constant.

All 462 IR/evaluator library tests pass. Seven source integration tests pass, including a reduced unconnected array-flow connector, all seven analytic channels in BDF/RK and Auto/Interpreter modes, and rejection of a forged zero-assignment prefix. OMC independently generates the same zero-flow equations and matches those seven analytic channels with maximum absolute error 8.197e-8. This follows MLS §9.2 and Appendix B.1a. An explicit written zero equation used an existing Direct path; the retained connector reduction exercises the originating missing seed.

All 512 core integration tests and affected all-target/all-feature Clippy pass. The fixed canary target/msl/multibody-zero-assignment-canary passes at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, dirty digest f6a825d032b0c4c0fbea179225d9380e9990a6125be82b0122b7710dee3be07b. All 20 phase/status and band rows match the preceding shared-interface canary: nine compared models, all 175 trajectory and initial channels high, and zero deviating, skipped, missing, excluded, or nonidentifiable comparisons.

The normal-budget origin still fails the 12-second Sim limit after successful initialization; Solve takes 7.107 seconds. A separate 60-second diagnostic at the source experiment’s rtol=atol=1e-8 completes five simulated seconds in 19.251 seconds of runtime. Its 967 shared OMC channels are all high, including initialization; the worst channel bounded normalized L1 error is 8.861e-6. This remains diagnostic evidence, with no normal-budget high-coverage credit. An earlier diagnostic mistakenly used rtol=atol=1e-6; its separate artifacts are retained and excluded from performance conclusions.

The new Solve artifact preserves every canonical continuous program and changes only the continuous refresh owners. Algebraic seed rows increase from 2,030 to 2,045 and numerical projection stages decrease from 51 to 36. The complete-seed staged guard is unchanged. The two coupled blocks still lack 74 seeds. One remaining residual is now mapped through OMC alias signs: Rumoca row 1426, after evaluating this model’s rotation parameters, is the negative of OMC regular residual 1716 and therefore has the same zero set. The typed call owner is a transpose followed by matrix multiplication; OMC chooses different tearing assignments. This establishes one equation mapping, not equivalence of the complete coupled systems.

Evidence is under .git/multibody-campaign/gyroscopic-sim-timeout/: zero-assignment-receipt.json, zero-assignment-canary-delta.json, zero-assignment-origin-receipt.json, zero-assignment-solve-trace-delta.json, zero-assignment-strict-omc-comparison.json, and coupled-torque-origin-omc.json. The subsequent full 566-model gate passes; its measurement and remaining scope are recorded at the top of this ledger.

Independent additive assignments

The next reduction exposes a numerical defect in the old AffineResidual certificate: the residual x + 4 - 13 with a starting guess of 1e30 returns an isolated value of zero, although its solution is nine. The former evaluation x - residual/coefficient loses the small independent terms. The RED test is additive-offset-red-1.log; this is a general equation-isolation defect rather than a model-specific exception.

Solve construction now issues Additive with a finite, nonzero constant target coefficient and weighted, target-independent source registers. Reverse producer order combines shared contributions before visiting their children, so repeated DAG expressions cannot expand into an exponential term tree. The original program prefix still executes; the isolated value uses only the independent register selection. Shape metadata is immutable and borrowed during evaluation. One IR-owned final scalar materializer serves both canonical refresh owners and the evaluator’s compiled schedules. The old residual-correction representation and its readers are removed; Solve schema 68 replays the new certificate. Complete seed coverage, coefficient singularity, and nonlinear branch guards remain required under SPEC_0036 / SPEC_0043 §6 and MLS Appendix B.1a.

The reduced AdditiveTorque.Probe retains the actual moment-equation form zeros(3) = torque_a + transpose(R)*torque_b - cross(r,force), with the rotation inside a typed function call. At R=[0,-1,0;1,0,0;0,0,1], r={1,2,3}, torque_b={x,2*x,3*x}, and force={2*x,3*x}, the exact result is torque_a={-3*x,3*x,-4*x}. All ten channels follow x=exp(-time) under BDF/RK with Auto/Interpreter, despite torque starts of 1e30. OMC’s flattened equation preserves the same rotation and cross product; regular assignments 13–15 isolate the independent torque expressions. Its twelve CSV rows agree with the ten analytic channels to at most 1.093e-7 absolute error.

Focused validation passes 466 IR/evaluator tests, all 514 core integration tests, and 429 solver control tests. The mode-dependent pivoting fixture now asserts its newly available additive y=3-x seed and explicitly verifies that incomplete seed coverage still requires the complete coupled solve; its singular-coefficient mode and final solution assertions are unchanged. Regressions cover the original cancellation, bounded sharing after forty doublings, prefix input errors, materialized execution, cancelled/nonlinear coefficients, and rejected wire changes to terms, coefficient, and prefix. The fixed 20-model canary passes with all phases and agreement bands unchanged: nine compared models, 175 trajectory and initialization channels high, and no missing, skipped, excluded, or deviating comparisons. Four models have small error-score changes, recorded in the delta; the band rows are not identical. The full 566-model gate then passes as reported above.

The originating GyroscopicEffects normal attempt still times out at 12 seconds, after successful initialization and 6.651 seconds of Solve. A diagnostic from freshly recompiled DAE, at rtol=atol=1e-8 with a 60-second runtime limit, finishes five simulated seconds in 18.024 seconds. All 967 shared OMC channels are high, including initialization; 1,309 Rumoca-only channels remain outside that comparison. This is diagnostic evidence, not normal-budget coverage or a controlled speedup measurement. An initial diagnostic accidentally consumed a DAE predating the fixed-attribute repair; it is preserved and excluded from current-source correctness/performance conclusions.

The fresh source export changes provenance identifiers because of source path spelling. Matching all 2,668 source files by canonical path and identical bytes allows the span IDs to be mapped explicitly. After that mapping, the only Solve changes are schema version and continuous refresh owners: the original equations, initialization, events, typed calls, and other products are identical. Algebraic seed rows increase from 2,045 to 2,127, numerical projection stages decrease from 36 to two, and missing seeds decrease from 108 to 26. Derivative seed rows grow from 554 to 602; its missing seeds decrease from 74 to 26. Both plans remain uncertified as complete causal solutions.

The next concrete residual is logical row 1399, program 426, output 2, matched to bodyCylinder1.body.frame_a.f[2] (Y871). Its target is inside cross(r_CM,frame_a.f)[3], with a runtime coefficient r_CM[1]. OMC’s regular moment assignment 1688 contains the same product; its alias map negates the body’s torque into bodyCylinder1.frameTranslation.frame_a.t[3]. OMC solves the force using a different translational equation (1686). Symbolic expansion of the canonical tensor program, after that XML torque alias and the two zero center-of-mass components, gives exactly the negative of OMC’s assignment residual; their sum is zero. The target coefficient is -r_CM[1]. A general tensor affine projection is a hypothesis for the remaining isolation gap; a zero or nonfinite coefficient must retain the original residual, and the complete-seed guard must remain. No compiler change for that next hypothesis has been made.

Artifacts are under .git/multibody-campaign/gyroscopic-sim-timeout/: additive-offset-libraries-2.log, additive-offset-core-1.log, and additive-torque-omc/ (source, flattened equations, backend info, CSV, analytic comparison). Current validation receipts include additive-assignment-canary-delta.json, additive-assignment-full-delta.json, additive-assignment-origin-receipt.json, additive-assignment-solve-delta.json, additive-assignment-source-provenance-delta.json, additive-assignment-diagnostic-configuration.json, and additive-assignment-current-omc-comparison.json. The original equation mapping remains in coupled-torque-origin-omc.json.

Parameter preparation during shape inference

An isolated FullRobot diagnostic takes 8.83 seconds in Flatten, then reaches its prior ToDae ED019 refusal. The diagnostic does not replace its timed cohort failure. CPU samples show parameter-map construction and copying; the sampled DWARF stacks are incomplete, so no inclusive CPU percentage is claimed. A debugger stop on the 32nd build_eval_context call establishes the path from Context::build_parameter_lookup through DimensionScope::value to a fresh ParamEvaluator for an individual shape query.

Reduced tests return the correct shapes but fail the work bound: a matrix walk prepares the inventory 32 times, and two comprehension branches prepare it six times. DimensionScope now lazily prepares one existing evaluator per immutable parameter inventory in a walk. Cloned lexical scopes share that preparation but retain separate index bindings; separate roots and function invocations prepare their own inventories. This caches no expression results. The three focused tests pass, including sibling scopes and index shadowing, as do 145 evaluator tests, 630 Flatten tests, all 505 compiler-core tests, and affected all-target/all-feature Clippy. Formatting and diff checks pass.

The seven-model origin run recovers each model’s earlier stage under the normal budgets: FullRobot, IMS_Start_Polyphase, and LightningSegmentedTransmissionLine reach their prior DAE refusals; PrismaticConstraint, PlanarFourbar, and PointGravityWithPointMasses2 reach Solve; GyroscopicEffects initializes before failing simulation. There are zero completed simulations or comparisons, so the focused run exits 1 with parity unmeasured. FullRobot and IMS still take 9.15 and 9.63 seconds in Flatten; the full cohort must test whether the recoveries hold under its load.

The fixed canary multibody-dimension-preparation-canary passes at dirty-tree digest 29302162fb394ef5bd448c433d4286a3ac32c2595d6af2e21b717ec8fafdc6ac. All 20 phases/statuses and channel-band rows are unchanged. Nine models compare high, with all 175 trajectory and initialization channels high and zero missing, skipped, nonidentifiable, or deviating comparisons. Tier 1 is complete. The preceding multibody-dimension-preparation-full measurement confirmed six of the seven focused recoveries; IMS_Start_Polyphase remained a Flatten timeout. All prior high-parity models and channel bands are retained. No new high-parity model is claimed.

Evidence is under .git/multibody-campaign/fullrobot-flatten-profile/, fullrobot-flatten-callstack/, and dimension-preparation-{red,green}-1.log. The origin and canary deltas are dimension-preparation-origin-delta.json and dimension-preparation-canary-delta.json. The governing ownership rules are SPEC_0029 §5 and SPEC_0033 §§2–3/6a; the existing MLS array and lexical-scope rules remain unchanged.

Borrowed inventories in the shared constant evaluator

The remaining IMS_Start_Polyphase regression exposes copying between separate shape walks. Its isolated diagnostic spends 8.51 seconds in Flatten; leaf CPU samples again identify parameter-map preparation. Eight reduced independent queries return the right shapes but materialize the 128-binding inventory eight times, failing the zero-copy work bound.

EvalEnvironment now supplies read-only values, dimensions, and functions to the existing constant interpreter. Owned and borrowed inventories share EvalContext’s single scoped lookup order. ParamEvaluator borrows the stable Flatten maps, prepares enumeration identities and function-name aliases, and materializes only scalar values actually read. It retains no expression-result cache and cannot outlive the borrowed inventory. Function definitions and shape vectors are borrowed, including through short function aliases.

The reduced tests compare owned and borrowed interpretation across sibling and root scopes, unknown values, enums, function formals, and shape-only metadata. Resource tests count actual owned-context parameter insertions and require zero for scalar-only shape queries. Pointer checks prove that function aliases and dimensions retain their original backing data. All 147 evaluator and 630 Flatten tests pass; all 506 compiler-core tests pass.

An identical source probe in Rust and OpenModelica a96aa1a-cmake preserves 2×2 and 5×2 arrays in sibling components, despite a different root-scope parameter. Both Rumoca solvers and OMC produce state slopes 1.5 and 12, with their sum at 13.5. OMC emits no warnings or errors; its six trace rows differ from those affine trajectories by at most 1.34e-15. This checks MLS §§5.3, 10.1, and 12.4; no compiler branch names an MSL model.

Evidence is .git/multibody-campaign/borrowed-parameter-receipt.json, borrowed-parameter-omc/comparison.json, and the corresponding test logs. All-target/all-feature Clippy passes for the evaluator, Flatten, and main crate. The final evaluator rerun passes all 147 tests, including the insertion and borrowing checks. The fixed canary passes with all 20 phase/status and channel-band rows unchanged: nine compared models, all 175 trajectory and initial channels high, and zero missing/skipped/nonidentifiable/deviating comparisons. Its durable delta is borrowed-parameter-canary-delta.json.

The full 566-model gate passes in 249.09 seconds with 493 Flatten completions, 286 DAE models, and 256 Solve models. IMS_Start_Polyphase recovers its earlier ToDae balance frontier. All 566 model bands and channel-band counts are unchanged from the preceding full run: 143 compared high, 16 tracked exclusions, zero missing/nonidentifiable/deviating comparisons. MultiBody remains at 40 Flatten, 33 balanced DAE, 17 initialized, and 14 strict-high models out of 42. The full receipt is borrowed-parameter-full-delta.json; no new high-parity model, baseline promotion, commit, or completed quick/full verification is claimed.

Boolean parameters determined at initialization

Fourbar_analytic’s ED020 refusal is traced through a debug build to ordinary initial-equation lowering, then arithmetic binary-expression construction. This excludes the competing function-builtin and record-field explanations. Flat initial row 0 subtracts a Boolean-valued selectBranch call from jointSSP.prismatic.positiveBranch. The row’s left-hand reference and the declared Boolean parameter carry the same occurrence identity; its source declaration has fixed=false in PrismaticWithLengthConstraint.mo.

A reduced source with no MultiBody dependency reproduces ED020 on branch = q > 0 in an initial equation. Another initial equation sets q=-2, despite a start guess of 5. OpenModelica accepts the identical source without warnings and produces q=-2, branch=false, and y=-time in all six rows. The correction must preserve MLS §8.6’s initialization semantics: selecting the branch from the start guess or converting Boolean equality into numeric subtraction would be incorrect. A typed initialization owner and regression now pass focused validation. The DAE owner preserves the exact unbound, non-Real fixed=false parameter, its type/shape, and source provenance. Wire replay and structural transformation reconstruct it through the checked API. Solve substitutes the definition into initialization residuals and commits its value after solving. Cyclic parameter definitions and initialization-determined String storage retain explicit unsupported diagnostics.

A stronger independent probe exposed a second issue: initialization grouped all rows sharing an unknown into one numerical block even when their matched dependencies were one-way. With a Boolean branch changing between the state guess and solution, the combined line search stalled. The planner now uses the shared dependency-first SCC decomposition to order matched rows, retaining mutually dependent rows together and all unmatched consistency checks.

The three reduced core tests pass with BDF and RK-like simulation. They cover opposing guesses and solved signs, Boolean/Integer dependency chains, a function call, reversed equation orientation, and a state crossing zero after initialization while its parameters stay constant. OpenModelica accepts the three identical positive source variants without warnings or errors; six rows per variant agree with the analytic trajectories within 1.78e-15. The complete compiler-core suite passes all 509 tests. The seven affected libraries pass 1,451 tests; four initial-owner integration tests pass. All 243 architecture checks, 17 repository gates, and affected all-target/all-feature Clippy pass. The schema pin now names template version 6, and the DAE module review below accounts for its new owner. No coverage claim was made from focused tests alone. The following complete measurement closes that validation. Details are in boolean-initial-parameter-receipt.json.

The fixed 20-model canary passes with every phase and band unchanged: nine compared models, 175 trajectory/initial channels, and no missing, skipped, or deviating comparisons. The three analytic origins now produce balanced DAEs. Fourbar_analytic reaches EL005 with 1,622 of 1,659 equations matched; Engine1b_analytic and PlanarLoops_analytic reach the 10-second Solve limit. That focused origin run has no simulations or comparisons and reports parity unmeasured, not success.

The full multibody-initial-parameter-full gate passes in 312.16 seconds at the commit and digest stated above. DAE completions increase from 286 to 289, and balanced DAEs from 274 to 277. All 566 model bands and channel-band counts are unchanged: 143 compared high, 16 existing exclusions, and zero missing, nonidentifiable, or deviating comparisons. No new high-parity model is claimed. GyroscopicEffects now hits the Solve limit before initialization; its preceding full run spent 8.65 seconds in Solve before reaching a 12-second simulation timeout. This regression remains visible and requires profiling. The complete delta is boolean-initial-parameter-full-delta.json. The speed gate passes, but combined verify quick and verify full remain outstanding.

Scalar-row owner lookup

The latest full run exposes GyroscopicEffects exceeding its unchanged 10-second Solve budget. It has no initialization-defined parameter owners. A separate diagnostic profile attributes 15.44% of leaf samples to DaeView::continuous_owner and 6.54% to continuous_owner_for_scalar_row. GDB stops inside that scan with the actual caller explicit_derivative_preferences during structural preparation. This proves the repeated lookup cost, but does not attribute the entire timeout to it. The profile is diagnostic evidence, not a successful normal-budget attempt.

The reduced resource test makes 256 row queries in reverse order and fails because the old lookup materializes 32,896 owner views. The checked equation constructor now derives each owner’s exclusive scalar-row end from its predecessor and checked row count. Overflow is rejected before insertion. The immutable owner sequence supports binary search, including duplicate ends for empty families. It retains one entry per semantic owner; it neither expands domains nor changes residuals, source order, or wire operations. SPEC_0032 §§1–2 and SPEC_0036’s construction/index rules govern this change.

Four focused tests pass: the resource bound, mixed scalar/family boundaries with multiple body equations and canonical JSON/binary replay, empty families, and compact domains with more than u32::MAX total scalar rows. The four affected libraries pass 479 tests; compiler core passes 509, architecture 243, and repository gates 17. All-target/all-feature Clippy passes after extracting a repeated-body test helper. Diagnostic compilation of the unchanged GyroscopicEffects source produces byte-identical DAE and Solve exports before and after this optimization; dae-equality.json and solve-equality.json retain the hashes. The first post-change diagnostic omitted the sibling ModelicaServices source root and stopped at name resolution; the corrected command loads the complete pinned library root. These exports are separate from normal-budget gate attempts. The fixed 20-model canary passes with all phase, model-band, and channel-band counts unchanged: nine compared models and 175 high trajectory/initial channels, with no skipped, missing, or deviating comparisons. The focused GyroscopicEffects origin still exceeds the 10-second Solve budget and has no parity measurement. This failure remains recorded.

The complete multibody-scalar-owner-full gate passes in 283.02 seconds at the commit and digest stated above. All 566 model bands and channel-band counts are unchanged: 143 compared high, 16 existing exclusions, and zero missing, nonidentifiable, or deviating comparisons. DAE completions remain 289, balanced DAEs 277, and successful simulations 159. Successful initializations recover from 175 to 176: GyroscopicEffects completes Solve in 7.74 seconds, with 7.68 seconds in lowering, then initializes and reaches its 12-second simulation timeout. This full-run result does not erase the focused timeout or establish reliable performance headroom. The speed gate passes; no new high-parity model, baseline promotion, or combined quick/full pass is claimed. Evidence is in .git/multibody-campaign/gyroscopic-solve-timeout/, including scalar-owner-canary-delta.json, scalar-owner-origin-delta.json, and scalar-owner-full-delta.json.

A separate diagnostic attaches sampling to the actual msl-fast worker only during Solve, using the worker’s one-thread compiler setting. Its 670 CPU samples put 13.43% of leaf cost in FunctionRebuilder::rebuild_function. Inspection finds a whole-expression-arena scan for orphaned scoped expressions inside each function reconstruction. This is the next resource-cost hypothesis, not a proven attribution of all sampled time: it needs a reduced RED test with orphaned scoped expressions and unrelated model expressions before changing the producer. The worker was stopped after Solve; this diagnostic has no coverage credit. Its receipt and analysis are under worker-profile/.

The SPEC_0043 module review measures 17,347 core + 5,431 wire = 22,778 lines. Against the initialization-owner inventory below, equations increases from 799 to 835 lines and model excluding wire from 8,373 to 8,376; every other group is unchanged. The 39-line increment retains checked overflow handling and immutable derived lookup metadata. The old linear scan is removed, and serialization still emits only the existing semantic operations. There is no second equation representation or compatibility reader to retain. Only the total acknowledgment crosses a 250-line step; review triggers are unchanged.

Initialization-owner module review

Under SPEC_0036’s DAE milestone review and SPEC_0043 §1, production DAE source is 17,308 core + 5,431 wire = 22,739 lines, compared with the preceding derivative-owner review’s 17,163 + 5,407 = 22,570. This adds 145 core and 24 wire lines; only the core acknowledgment crosses a 250-line step.

The new 128-line model/initial_parameters.rs contains the checked constructor and immutable view. Remaining core additions register the arena, identity, error, count, and freezing. The owner cannot be an ordinary parameter binding: its value may depend on initialization unknowns. It cannot reuse the discrete initial-value owner, whose reads must already be settled. Wire replay shares InitialValueWire with discrete values, replacing the old discrete-only record name, but retains separate checked construction and ordinal checks. The wire accepts only the current schema. Review found no obsolete compatibility reader, duplicate parameter-binding path, or removable semantic check.

The evidence is .git/multibody-campaign/boolean-initial-equation-callstack/ (run-debug.log, initial-branch-flat.json, row-identity.json) and boolean-initial-equation-probe/ (Source.mo, the Flat export, ED020 diagnostic, and comparison.json). The initial optimized-worker debugger attempt lacked source symbols; the recorded stack comes from the subsequently rebuilt debug worker. Engine1b_analytic and PlanarLoops_analytic share the error category; their exact rows have not yet been independently traced.

Scope and initial census

The source is MSL 4.1.0, with ModelicaTest pinned at 8ae3d35c24e519cb2996cab20f3b13daf2b0c50a. An OpenModelica semantic inventory (getClassNames, getClassRestriction, isPartial) identified:

• 42 MultiBody root examples;
• 111 non-partial models under ModelicaTest.MultiBody;
• 171 non-partial models in the MultiBody library, including components that require an enclosing model and connections.

The library also contains functions, connectors, types, and partial classes. The inventory is a scope record, not evidence that those declarations are covered. Component and function coverage still needs to be mapped to reviewed test embeddings; the 42 examples alone do not establish complete library coverage.

The initial 42-example diagnostic run at the branch base used the normal 10-second phase budgets and 12-second simulation budget with a 14-second parent watchdog. OpenModelica a96aa1a-cmake simulated every selected example. Rumoca reported 20 balance failures, six structural-analysis failures, six timeouts, five DAE-construction failures, one runtime-contract failure, and one instantiation failure. The remaining three simulations were compared; none had a deviating channel, skipped comparison, or missing trace.

Artifacts are retained locally under target/msl/multibody-baseline, with provenance and hashes under .git/multibody-campaign. This focused diagnostic does not supply a full 566-model cohort parity claim or baseline promotion.

Array constructor equation dimensions

The originating model is Modelica.Mechanics.MultiBody.Examples.Elementary.SpringMassSystem. Its two Joints.Prismatic instances contain this torque balance:

zeros(3) = frame_a.t + frame_b.t + cross(e*s, frame_b.f);

The expected equation count is three per joint. At the branch base, Flat equation rows 10 and 57 each recorded scalar_count = 1. This explains exactly the observed deficit: 1465 equations for 1469 unknowns.

The first divergence is in flattening. The array-only dimension context does not establish the scalar shape of s, so the right-hand cross product cannot prove its shape there. The flattener also omitted the constructor shape on the left. Connection-edge removal is not the cause: the model’s oc_break_edge_scalar_count is zero. Inspection of the Flat artifact found no mismatched cardinalities for direct variable-reference left sides.

The semantic basis is MLS 3.6 §§10.3.3, 10.3.5, 10.6.1, and 10.7; the phase owner follows SPEC_0007 and identity handling follows SPEC_0001. The fix recognizes constructor calls through their exact predefined declaration IDs and carries their complete dimensions into equation cardinality and the Flat tensor domain. fill retains the dimensions of an array-valued element after its new axes. Unknown or invalid extents do not establish a constructor shape. DAE balance checks are unchanged.

The regression scaled_cross_product_preserves_all_three_torque_equations reduces the defect to seven scalar unknowns. Before the fix it fails with five equations. Its independent runtime oracle is torque(t) = {4*(2+t), -3*(2+t), 0}. Additional focused checks cover matrix and higher-rank constructors, empty extents, structural dimensions, unknown dimensions, same-spelling non-predefined declarations, and the scalar result of a vector dot product.

The repaired real-model Flat artifact changes only the two equation cardinalities and their structured domains. Each new domain has one binder over 1:3; the original equations, declaration identities, and spans remain intact.

Tier 1 validation:

• All 624 flattening tests pass.
• The end-to-end torque regression passes its equation-balance and analytic trace checks.
• Clippy passes for rumoca-phase-flatten and the rumoca library and core integration suite, with all features and warnings denied.
• The fixed canary has no model-band or failure-category transitions. All eight completed simulations are compared with no deviating channels, skipped comparisons, or missing traces. The other twelve targets retain their failures; the command’s successful exit does not make them supported.

The canary’s before evidence is the same twenty members of the complete Tier 2 run at 82d87a16d11f6c6e86069510cfec471874da480a, whose source tree 0436b778ec9a36a53a8d935fe29a98a89344edec is identical to the branch base. After evidence is in target/msl/multibody-constructor-canary; the per-model delta and hashes are in .git/multibody-campaign/constructor-canary-delta.json. This is a Tier 1 regression check, not a cohort parity claim.

The follow-up 42-example diagnostic removes 19 of the 20 balance failures. Constraints.PrismaticConstraint still has a two-equation deficit, down from sixteen. SpringMassSystem now has 1469 equations and 1469 unknowns, but structural matching still fails (1394 matches for a 1406-by-1406 system). These remaining failures are not exceptions or passing models.

Rotational3DEffects.ActuatedDrive newly completes simulation. All 455 compared channels, including initialization, agree with OpenModelica within the high band. Every completed simulation in this focused run is compared; there are four comparisons, no skipped or missing traces, and no deviating channels. Per-model changes and artifact hashes are retained in .git/multibody-campaign/constructor-multibody-delta.json.

Manual review of ActuatedDrive identifies two equivalent inertias of 2, driven by sin(2*pi*time) with zero initial angle and angular velocity and zero gravity. An independent check of all 501 Rumoca samples verifies both revolute and rotor1D against w=(1-cos(2*pi*t))/(4*pi) and phi=t/(4*pi)-sin(2*pi*t)/(8*pi^2). Maximum absolute errors are below 1.3e-6 for velocity and 7.7e-7 for angle, with exact zero initialization. The check and trace hash are in .git/multibody-campaign/actuated-drive-analytic.json. This reviewed case exercises the new constructor-shape path together with the existing revolute-joint, rigid-body, one-dimensional rotor, mounting, connection, initialization, and integration paths; it does not establish unrelated joint or state-selection behavior.

The complete 566-model Tier 2 run at 3dec54b64813bfb4e50d6177ab6b664cccb59ef4 passes. It compares 130 models, all strict-high, with zero deviating channels or missing traces; 17 unchanged reviewed policy exclusions remain skipped and do not count as supported. The only model-band transition from the preceding full run is ActuatedDrive, absent to high. Evidence is retained in target/msl/multibody-constructor-full and .git/multibody-campaign/constructor-full-receipt.json. The tracked source stayed unchanged throughout the run; the local untracked FastDyn communication file accounts for the comparator’s dirty-worktree flag. Release gates have not yet been rerun for this branch.

Declared function result dimensions

The remaining balance failure after the constructor fix was Constraints.PrismaticConstraint: 2427 equations for 2429 unknowns. Its freeMotionScalarInit.initAngularVelocity component contains:

Frames.angularVelocity2(R_b) =
  Frames.resolve2(R_b, Frames.angularVelocity1(R_a)) + w_rel_b;

Frames.angularVelocity2 declares the output w[3], but Flat row 160 counted this equation as one scalar. The precollected executable function already retained the resolved declaration identity and fixed output shape; equation shape inference did not consume that evidence. Connection-edge removal again contributes zero and cannot explain the deficit.

MLS §§12.4.3 and 10.6.1 establish the first output and equation dimensions. MLS §12.4.6 permits automatic vectorization only for a function with one scalar result. A fixed array result therefore establishes its declaration’s axes; a scalar result alone cannot establish the call’s dimensions. SPEC_0007 assigns this work to flattening, using SPEC_0001 declaration identity.

The flattener now indexes declaration-proven array results by DefId before flattening equations. Every exposure of a declaration must agree on the shape. Conflicting exposures, deferred extents, and scalar results remain unknown at this boundary. Equation cardinality and the structured domain consume the same shape. The real-model Flat difference is exactly row 160’s cardinality, one to three, and its corresponding 1:3 domain.

The end-to-end angular-velocity regression fails before the fix with seven equations for nine unknowns. After the fix it balances and follows the independent analytic trace {3*time, 4*time, 5*time}. A scalar-function vectorization control verifies all three components of shift(x)={1,2,3}. Unit controls distinguish same-spelling declarations, refuse conflicting exposures in either order, and refuse deferred dimensions even when provisional effective dimensions look fixed.

Tier 1 validation passes: all 628 flattening tests, the angular-velocity and vectorization regressions, the previous torque regression, formatting, and focused all-feature Clippy with warnings denied. The fixed twenty-member canary has no band, phase, or failure-category changes relative to those same members in the complete run at 3dec54b6. All eight completed traces are compared with zero skipped, missing, or deviating channels; twelve targets retain their failures. Evidence is in target/msl/multibody-function-shape-canary and .git/multibody-campaign/function-shape-canary-delta.json.

The same 42-example diagnostic now proves PrismaticConstraint balanced at 2429 equations and 2429 unknowns. It remains structurally singular, with 2287 matches in a 2313-by-2313 system, and is not supported. The other 41 models retain their phase and failure classifications. All four completed traces are compared, with zero skipped, missing, or deviating channels. There are no remaining balance failures in this focused set. The remaining 38 failures are 18 structural-analysis failures, twelve timeouts, six DAE-construction failures, one runtime-contract failure, and one instantiation failure. Evidence is in target/msl/multibody-function-shape-after and .git/multibody-campaign/function-shape-multibody-delta.json; producer hashes are in .git/multibody-campaign/function-shape-proof.json.

The complete 566-model Tier 2 sweep at 3bc697611242f10c19d93dcc3e93eb26d55172aa passes: 130 models compared, all strict-high, 9659 channels compared with zero deviations, and zero missing traces. The same 17 reviewed policy exclusions remain skipped and unsupported. No model changes agreement band relative to the complete run at 3dec54b6. The tracked source remained unchanged throughout this sweep. Evidence is in target/msl/multibody-function-shape-full and .git/multibody-campaign/function-shape-full-receipt.json.

State definitions through exact coordinate aliases

SpringMassSystem reaches a balanced DAE after the shape fixes but originally matches only 1394 of 1406 structural rows. The direct state-definition search accepts a state on the left of a kinematic equation, but the model gives its kinematics through p1.frame_b.r_0, an algebraic coordinate exactly equal to body1.frame_a.r_0. The unique causal-definition map cannot supply this fact: the geometry equation and the alias equation both define that coordinate.

The structural change consumes the existing exact value-equality anchor, preserving its sign, before applying the existing state-demotion checks. It accepts either equation orientation and reconstructs a checked DAE. This is SPEC_0007/SPEC_0040 STRUCT-T04, with MLS §§8.3.1, 8.6 and der semantics; it does not implement general state selection or alias elimination.

The real model now matches all 1406 rows. Its 644 variables, 445 continuous equation owners and source map keep their original order and provenance. Six body-coordinate arrays become algebraics; the four fixed joint initial values remain 0.1, 0, 0.1, 0. At this step runtime still exceeds the 12-second simulation budget. The subsequent tensor assignment repairs below let the same model complete in 2.134 seconds with every compared channel strict-high in target/msl/multibody-gravity-complete-examples. An independent analytic check of both prismatic joints uses s = 0.1 + g/30*(1-cos(sqrt(30)*time)), with g = 9.80665. Maximum displacement, velocity, and acceleration errors over all 501 samples are 1.94e-5, 1.08e-4, and 5.79e-4, respectively, under the model’s original 1e-6 tolerance.

Three analytic kinematic regressions cover vector coordinates, a signed alias, and reversed equation orientation, including the initial displacement and velocity. All 141 structural tests and 13 initial-value alias-transfer runtime tests pass. Five holonomic-specific fixtures now declare StateSelect.always to continue exercising their intended reduction; their proof assertions are unchanged. Focused structural/core Clippy and formatting passed.

The diagnostic canary in target/msl/multibody-alias-state-canary adds FirstGrounded with every compared channel high and leaves its other members unchanged. The subsequent 42-example run at target/msl/multibody-alias-state-after exposes the counterexample below. These are focused working-tree measurements, not complete cohort claims. The state-demotion change was held until that counterexample was closed.

Receiver function redeclaration counterexample

The originating model is Modelica.Mechanics.MultiBody.Examples.Elementary.UserDefinedGravityField. The state-demotion work enables it to complete, but the comparison reports 48 deviating channels among 193, including five at initialization. Rumoca keeps gravity and the pendulum motion zero; OpenModelica’s initial vertical gravity is -9.780263581798753. This paused unrelated capability work until the closure recorded below.

The source redeclares world.gravityAcceleration to theoreticalNormalGravityWGS84(phi=geodeticLatitude). Before the receiver fix, Flat equation 27 still calls the default standardGravityAcceleration implementation with world.gravityType=NoGravity. That explains zero gravity without invoking a solver, initialization, or mechanical-reduction hypothesis.

The minimal receiver regression replaces a default Double with Triple. Instantiate records the correct slot and implementation DefIds, but Flatten selects Double through direct and inner/outer receivers. The first divergent producer is Flatten’s function override selection: it only consults the caller’s modification environment. MLS §§5.4, 7.2.2 and 7.3, SPEC_0001 and SPEC_0007 Stage 2 govern the repair.

The working repair supplies the receiver’s instance modification environment and retains the enclosing instance scope of bound modifier arguments. Three Flat identity regressions pass. Two runtime regressions check direct and inner/outer calls across every sample, including nested instances with distinct parameter values. The nested test caught a second defect in the intermediate repair: both instances used the declaration’s default value. Carrying the modifier’s enclosing instance identity and path closes that focused regression. The complete Flatten test suite passes (687 tests across its unit, integration, and documentation groups), as does Flatten’s all-feature, all-target Clippy check. The originating model’s Flat output now selects only the WGS84 implementation and carries the latitude argument from the correct instance.

The normal 12-second run in target/msl/multibody-receiver-gravity and a separate 60-second diagnostic run both reach consistent initialization, then time out without a complete trace. The extended diagnostic changes no supported coverage count or benchmark budget. Public live stepping through 0.5 seconds produces nonzero pendulum motion and the expected initial gravity, but it is not a replacement for the complete OMC comparison. Algebraic output observation is expensive: at 0.02 and 0.04 seconds it takes 26 and 32 full projection sweeps respectively. Diagnostic tracing identifies two unsettled singleton blocks; the exact producer and assignment certificates are under investigation. The missing proof was in Solve’s exact-assignment constructor: scalar additive residuals admitted an isolator, but the equivalent TensorBinary sum did not. The repair projects the operands of the requested tensor element through the existing stride/lane checks and applies the same additive proof. It retains the compact program and rejects cancelled targets and unisolated dependencies. Both positive tests failed before this change; all five tensor shape tests now pass, including stride identity and negative controls.

That repair reduced the full simulation to 1.69 seconds, but the comparison in target/msl/multibody-gravity-tensor-isolators still found five torque-channel deviations. The generated assignment builder allocated new registers after the largest destination start, overwriting live tensor lanes. A standalone materialization regression reproduced -1 instead of 4. The builder now uses the existing checked register-flow extent. Its single-output and grouped assignment regression passes; the Solve IR, evaluator, and solver unit suites pass all 823 tests.

target/msl/multibody-gravity-register-ranges completes the originating model under the unchanged normal budget: 193/193 channels high, all 193 initial channels high, zero missing, skipped, excluded, or deviating channels. Runtime is 1.690 seconds and build plus runtime is 3.396 seconds. This closes the originating counterexample. An independent two-state pendulum calculation uses phi_dot = w and w_dot = (-10000*g(20 + 10*sin(phi))*cos(phi) - 0.1*w)/100000.001, with the source WGS84 gravity function. SciPy DOP853 at rtol=1e-12, atol=1e-14 agrees at all 501 samples: maximum angle, speed, and acceleration errors are below 5.88e-7; gravity error is below 9.81e-12; all eight checked transverse torque channels are exactly zero. The fixed 20-model canary in target/msl/multibody-gravity-complete-canary retains nine compared models, all strict-high, with zero missing, skipped, excluded, or deviating comparisons. Every member’s compilation and simulation status is unchanged from the earlier state-alias canary. The additional repeated-prefix receiver regression passes, bringing the receiver identity suite to four tests. The full-cohort milestone remains pending. The affected packages pass all-target, all-feature Clippy.

Moving drive angular-acceleration counterexample (closed)

The combined 42-example validation in target/msl/multibody-gravity-complete-examples completes ten models. All ten have reference comparisons, with zero missing, skipped, or excluded models. Nine are strict-high. MovingActuatedDrive has one deviating channel among 624: bodyCylinder.body.z_a[3]; all 624 initial channels are high. The other 32 examples still fail before producing a complete trace. This counterexample paused unrelated capability work and the milestone commit until closure.

The model’s angular velocity satisfies w_a[3] = -r1.w*sin(revolute.phi), and its source declares z_a = der(w_a). Rumoca’s acceleration instead satisfies z_a[3] = r1.a*sin(revolute.phi) - r1.w*cos(revolute.phi)*revolute.w to 8.89e-16 over the complete candidate trace. The first term has the wrong sign under the product rule. Before structural reduction, DAE owner 401 retains z_a = der(w_a). After reduction its product-rule expansion puts (r1.a*r1.e)*revolute.R_rel.T where the source requires revolute.R_rel.T*(r1.a*r1.e). Structural differentiation treated multiplication as commutative in the second term of the product rule. Both forms typecheck for a square matrix and vector, so shape checks alone cannot establish correctness.

Three polynomial regressions reproduce wrong derivatives for matrix–vector, vector–matrix, and matrix–matrix products. The repair preserves both operand positions in the first- and second-order product rules and tests the value plus both derivatives over every sample. These fixtures construct dynamic matrices from fixed parameter matrices and time coefficients, within the existing differentiability rules; no unsupported operation is admitted to make the test pass. All 419 core integration tests pass, including these three regressions. This follows SPEC_0007’s structural transformation contract and MLS §3.7.2 (der) and §10.6.4 (ordered vector and matrix multiplication).

The originating run in target/msl/multibody-moving-drive-product-order compares all 624 channels as strict-high, including all 624 initial channels, with zero missing, skipped, excluded, or deviating comparisons. Runtime is 3.220 seconds under the unchanged normal budget. Its corrected acceleration satisfies z_a[3] = -r1.a*sin(phi) - r1.w*cos(phi)*revolute.w to 8.89e-16 over all 501 candidate samples. This closes the originating counterexample; the regenerated structural DAE retains owner 401 and residual 4450, while product node 4446 now has the required matrix on the left. All 141 structural tests and the affected packages’ all-target, all-feature Clippy checks pass. The fixed canary in target/msl/multibody-product-order-canary retains nine compared models, all strict-high, with zero missing, skipped, excluded, or deviating comparisons. All 20 members’ stage and simulation statuses are unchanged from multibody-gravity-complete-canary. The full example-set regression run in target/msl/multibody-product-order-examples compares all ten completed models as strict-high: 3759 channels, including initialization, with zero missing, skipped, excluded, or deviating comparisons. All 42 stage and simulation statuses are unchanged from the preceding example run; the angular-acceleration channel changes from deviating to high. The other 32 examples remain failures. The named-commit full 566-model milestone is next; these focused runs do not establish a cohort claim or baseline promotion.

The full 566-model gate in target/msl/multibody-state-tensor-full passed at 4716a8291f26302853aef2010b97b5e8d84e4fa2 (tree 5161f078ed87db562a551d149c351b83cc660da9). It compares 138 models, all in the high trajectory band, with 12126 channels and zero deviating channels or missing traces. Sixteen completed models retain reviewed comparator exclusions. Every previously high model remains high. Eight additional models enter that band: MultiBody DoublePendulum, ForceAndTorque, Pendulum, SpringMassSystem, UserDefinedGravityField, MovingActuatedDrive, plus Rotational FirstGrounded and Translational Sensors.

This result needs two explicit qualifications before further breadth work:

• DCPM_Drive, previously completed under a reviewed comparator exclusion, now fails Solve lowering: continuous algebraic row 133 cannot substitute der(dcpm1.airGapDC.flange.phi) because its matched state residual is not a subtraction. Its disappearance from the completed set explains the exclusion count changing from 17 to 16; the policy file is unchanged. This is a new execution regression to repair, not improved comparator coverage.
• Clocked.Examples.Elementary.RealSignals.AssignClockToTriggerHold retains an existing near initial channel: triggeredSampler.y is 0 while OMC gives 0.1 after the first clock tick at time zero. The error persists until the next tick at 0.02 seconds. Its trajectory score still falls in the high band, but that aggregate must not hide a potential clock/event semantic defect. Rumoca already samples the sine and toggles/holds the Boolean at time zero; the triggered non-clocked sampler fails to respond. The following repair closes this counterexample before resuming the MultiBody compile frontier.

The full run takes 272.260 seconds. Source files were unchanged during the sweep; artifact digests and exact counts are recorded in .git/multibody-campaign/state-tensor-full-receipt.json. No baseline is promoted.

Complete event iteration after the first clock tick

A minimal held-Boolean-clock fixture reproduces the missing sample at time zero. The same fixture with a first tick shifted to 0.01 seconds passes. Both include an independent initialization counter that must remain one, and their analytic sample oracle preserves exact event timestamps rather than moving a nearby sample across a tick.

DAE retains the hold, change, and Boolean when semantics in its typed temporal owners. Solve scalar row 2 computes trigger = held != pre(held) with a Fixed pre-read policy; guarded assignment owner 0 correctly reads the current input on the trigger’s rising edge and otherwise holds the sampler output. The first wrong operation is the runtime’s PostInitialClockTick filter: it excludes the Fixed trigger equation even after the held clock value changes. A pre-read policy does not identify an initialization-only equation.

The runtime now uses the complete SOLVE-C22 event pass after initialization clears for the first clock tick, as required by MLS §§16.3, 16.5.1 and 8.3.5. Initialization still defers clock-owned equations, and the existing event history commit and once-per-tick clock execution remain authoritative. The obsolete filter is removed. Both focused tests pass, along with all 421 core integration tests and all 417 solver unit tests.

The originating comparison in target/msl/multibody-clock-trigger-repair is high on all eight channels, including all eight initial channels with zero initial error. There are no missing, skipped, excluded, or deviating comparisons. This closes the clock-trigger counterexample; its fixed canary in target/msl/multibody-clock-trigger-canary retains all nine compared models as strict-high, with zero missing, skipped, excluded, or deviating comparisons. All 20 stage and simulation statuses are unchanged. Formatting and the affected packages’ all-target, all-feature Clippy checks pass. The DCPM_Drive execution-regression repair remains pending.

Signed derivative residuals

The DCPM_Drive failure comes from its matched state equation 491. Prepared DAE residual 4335 is +(der(dcpm1.airGapDC.flange.phi) - dcpm1.inertiaRotor.w): a unary plus around subtractive residual 4334. The state variable is 278 and its right-hand side is state 235. This is a valid equation under MLS Appendix B.1; neither the DAE nor structural reconstruction has lost its derivative definition. Solve’s derivative_rhs requires a subtraction at the root and incorrectly rejects the unary wrapper when algebraic row 133 needs that derivative.

Solve lowering now reads through unary plus and negation around a complete zero residual before applying its existing derivative-isolation checks. This preserves the equation’s solution and keeps executable derivative analysis in the SPEC_0007 Solve owner. Three checked-DAE regressions cover positive, negative, and nested wrappers. All fail with the original refusal before the repair. Afterward they verify both the derivative and its algebraic use at four assignments, including one outside the solution manifold. All 113 Solve tests and all 421 core integration tests pass, as do formatting and the affected packages’ all-target, all-feature Clippy checks.

The originating run in target/msl/multibody-derivative-wrapper-repair restores DCPM_Drive completion in 11.839 seconds, including 6.597 seconds of simulation preparation and 5.224 seconds of integration. The unchanged reviewed comparator exclusion still applies. A separate diagnostic using the same production comparator and fresh OMC reference checks all 590 common channels: 586 are high and all 590 initial values agree exactly. The four non-high channels remain precisely the previously reviewed idealDcDc.feedback.y and idealDcDc.powerController.u aliases in the two inverters. Against this reference, the prior candidate has four near channels; the restored candidate has two near and two deviating channels, all inside that existing integration-residual boundary. No other channel becomes non-high. The exclusion is unchanged and this model does not count as strict-high.

The clock model in that same originating run remains high on all eight channels with exact initial values. The fixed canary in target/msl/multibody-derivative-wrapper-canary retains nine compared models, all strict-high, with zero missing, skipped, excluded, or deviating comparisons. All 175 initial channels are high and all 20 stage and simulation statuses are unchanged from multibody-clock-trigger-canary. The execution regression is closed; the complete 566-model milestone follows at the repair commit. No baseline is promoted from these focused results.

Reference generation and cache evidence repair

The complete 566-model run at 7069c44e in target/msl/multibody-regression-restored-full failed its quality gate during reference-cache persistence. It supplies no accepted cohort parity number. Ten OMC attempts had been labeled successful despite an empty resultFile and no trace. Their runtime messages reported execution failures, but the reference producer discarded the wording Simulation execution failed. This was a producer error; the cache correctly rejected the incomplete success records. No candidate comparison was missing, and no compiler simulation status regressed from the preceding full run; DCPM_Drive returned to sim_ok under its unchanged reviewed exclusion.

The first reproduction, Modelica.Electrical.Digital.Examples.RAM, also identified an earlier reference setup error. Explicitly loading the generic MSL ModelicaServices made its resource loader pass modelica:// URIs to fullPathName, so initialization could not read the memory data file. The identical pinned model succeeds with OMC’s tool-specific services. Reference generation now lets OMC supply those services, requires a result file before reporting success, and retains runtime failure diagnostics. An independent warning-level assertion model confirms that warnings which permit successful simulation remain nonfatal. A cache-policy fingerprint invalidates references produced under the previous setup.

Regenerating JSON now replaces the destination after writing the complete payload, preserving historical traces which share a cache hard link. The empty-result, runtime-assertion, warning, and hard-link regressions each exercise their observed failure boundary. All 168 tooling unit tests pass, as do formatting and all-target, all-feature tooling Clippy.

The originating harness run in target/msl/multibody-omc-reference-origin produces successful OMC references for RAM, readRealParameterModel, and the MultiBody Pendulum. Only Pendulum currently completes in Rumoca; its 144 compared channels, including initial values, are high, with no missing or skipped comparison. This does not claim Rumoca support for the two resource examples. The final fixed canary in target/msl/multibody-omc-reference-final-canary retains all 20 stage and simulation statuses, nine strict-high comparisons, 175 high initial channels, and zero missing, skipped, or deviating comparisons. All 19 recorded historical reference-file hashes remain unchanged. The complete 566-model milestone follows at the repair commit; no baseline is promoted.

Complete milestone after reference repair

The complete 566-model gate at commit 13f0f133fac8edc80b7f55b53d820259bad74347 passed in 393.62 seconds. Its artifacts are in target/msl/multibody-reference-repaired-full, with the source audit and artifact hashes recorded in .git/multibody-campaign/reference-repaired-full-receipt.json. All 138 compared models are strict-high; 17 other completed models retain their reviewed exclusions, and zero candidate comparisons are missing. All 12,126 compared initial channels are high. Every previously strict-high model, including all 129 baseline-certified models, remains strict-high.

The run compiles 286 models, balances 274, attempts 267 simulations, and completes 155. Six attempts time out and 106 fail in the solver path. These stage classifications are unchanged from the preceding full run; the clock and DCPM repairs remain effective. OMC supplies successful references for all 155 completed candidates. The other 411 models remain visible in the full roster without an OMC simulation attempt in this run. The performance gate also passes with its complete 145-model cohort. The tracked source tree matches the named commit; the recorded dirty flag comes from the user-owned, untracked comm_fastdyn.md file. No baseline was promoted. These results do not establish complete MultiBody coverage.

Record fields in retained state constraints

Modelica.Mechanics.MultiBody.Examples.Constraints.PrismaticConstraint previously stopped in checked index-reduction reconstruction with a missing state-only substitution for expression 5431. This identity belongs to the reconstructed DAE at that reduction round, not the original DAE. It projects the orientation matrix from Frames.absoluteRotation into the retained position constraint from Joints.Prismatic:

frame_b.r_0 = frame_a.r_0 + Frames.resolve1(frame_a.R, e*s);

The differentiability analysis already understood this field projection. The first missing operation was exact value materialization in the structural phase, whose matching holonomic preflight also omitted record fields. Both now use the existing checked projection resolver and preserve nested function argument bindings. Unsupported projections still fail at the same typed boundary. The semantic basis is MLS 3.6 §§12.4, 12.6, and Appendix B; the owner follows SPEC_0007 and STRUCT-T03/T04 in SPEC_0040.

Three checked-DAE regressions reproduce the old rejection. They exercise both matrix fields of a record returned through nested calls with permuted arguments. Independent numeric expectations cover four parameter pairs, including negative, zero, and fractional values. The retained equation still relates the remaining matrix state, passes the state-only invariant, and has a computable structural system. All 144 structural tests, 421 core integration tests, formatting, and all-target, all-feature structural Clippy pass.

The production diagnostic inspection of the original Prismatic DAE now completes the exact formerly rejected state-296/RHS-5573 substitution, reducing the unmatched residue from 38 to 32. Subsequent exploratory reductions reach residue 20 before exhausting admissible candidates. The incomplete transformed system is not published; the model still reports its original structural singularity, 2287 matches for 2313 equations and unknowns. This repairs a construction boundary and adds no supported model. Before/after identities and artifact hashes are retained in .git/multibody-campaign/prismatic-manifold-triage.json.

The origin run in target/msl/multibody-prismatic-field-origin retains high parity for Pendulum and MovingActuatedDrive: two comparisons, all 768 channels and initial values high, zero missing or skipped comparisons. Prismatic’s OMC reference succeeds while Rumoca still refuses it. The fixed canary in target/msl/multibody-prismatic-field-canary has no changes to any of its 20 phase or simulation statuses relative to multibody-omc-reference-final-canary. All nine compared models and 175 initial channels remain high, with zero missing, skipped, or deviating comparisons. The recorded Tier 1 delta is .git/multibody-campaign/prismatic-field-canary-delta.json.

Exact scalar incidence in initialization

The next shared structural blocker is the initial-manifold guard: the GyroscopicEffects example reaches a holonomic constraint incident on the source-fixed revolute.phi. Its initial value is present in the DAE. The earlier preflight rejects the candidate because the existing post-initialization manifold projection may move that state; the later transferred-pin check is not the originating failure. The guard remains intact pending a joint initialization owner. MLS 3.6 §8.6 and Appendix B, SPEC_0007, and the SPEC_0036 draft / SPEC_0043 §4 describe the required ownership.

A prerequisite repair makes initialization incidence use the exact scalar projection already used by structural analysis. Matrix coordinates, structured domain points, substituted parameter bindings, and derivative definitions now retain the same scalar identity as their emitted residual programs. A transferred pin determines only its actual scalar. The prior whole-variable walk rejected three analytic matrix-initialization fixtures before simulation.

Five end-to-end regressions cover direct and structured matrix initial equations, derivative initial equations, permuted array parameter bindings, and inconsistent fixed matrix initial data. Both BDF and RK paths are checked against independent exponential trajectories or the required initialization refusal. All 426 core integration tests and 113 Solve tests pass, along with focused all-target, all-feature Clippy. The fixed canary in target/msl/multibody-initial-tensor-canary retains all 20 phase and simulation outcomes relative to multibody-prismatic-field-canary: nine comparisons and all 175 initial channels high, zero missing, skipped, or deviating comparisons. The artifact hashes and delta are in .git/multibody-campaign/initial-tensor-canary-delta.json. This is prerequisite coverage; it does not claim that the initial-manifold blocker is resolved.

Joint initialization and retained constraints

GyroscopicEffects first reaches the old guard at continuous owner 363, residual 6775, incident on state coordinates 474 and 502. Coordinate 474 is revolute.phi; its declaration carries fixed = true and a start expression. The source span is 2576..2588 in source 13533509827142484477. This identifies the admission problem before reconstruction, independently of the later initial-value preservation checker, which remains intact.

The replacement lowering includes every retained position and velocity constraint in the initialization residual, using the same scalar incidence as the emitted programs. Source starts proved independent of initialization unknowns supply given coordinates. Dependent starts and transferred pins remain equations alongside the unknown parameters they read. An opaque initialization owner derives row targets and unknown inventories from its checked plan and rejects overlapping projection, update, or given ownership. Wire schema 62 replays that constructor; the old independent inventories are removed. Root assembly checks the retained manifold count and storage bounds.

After initialization settles, manifold certification receives read-only state storage, including at the initial event boundary. Continuous-time correction retains its existing state projection. The early blanket admission guard is removed only with this replacement; inconsistent initial conditions remain errors.

Five end-to-end cases check uniform circular motion against independent analytic trajectories on both BDF and RK, partially fixed positions and velocities, and a fixed start that depends on a parameter solved during initialization. They also require rejection of inconsistent circle, time, and invariant constraints. All 248 Solve IR, 113 Solve lowering, 144 structural, and 417 runtime tests pass with the final given-coordinate handling, as do all 430 core integration tests. Focused all-target, all-feature Clippy passes across the changed compiler/runtime crates.

Lowering every fixed start to a numerical residual initially regressed ordinary FMI C export: Real x(start=2,fixed=true); der(x)=-x; was rejected by the existing parameter-only C initialization profile. Given-coordinate ownership restores this export without widening that profile. The packaged tensor decay conformance test now uses fixed starts and checks that legal FMI start overrides survive initialization. FMI 2.0.5 and 3.0.2 schema/VDM checks, source compilation, direct lifecycle checks, and FMPy ME/CS analytic traces all pass. The receipt is .git/multibody-campaign/joint-given-fmi-conformance-4.log; earlier attempts without the configured tools are not counted as validation. The combined fixed-canary validation is recorded below; this entry does not claim a MultiBody coverage gain.

GyroscopicEffects: forwarding-function equalities

The one-model run target/msl/multibody-joint-initial-gyroscopic gets past the initial-value admission guard, but fails structural analysis before numerical initialization. The frontend is balanced at 2276 scalar equations and unknowns. The structural reducer starts with 37 unmatched equations and unknowns, reaches 15 of each, then transactionally returns its original singularity error. The remaining equation owners are generated frame orientation and angular-velocity connections, rather than integration errors.

OpenModelica’s successful reference in target/msl/multibody-regression-restored-full uses the same MSL 4.1.0 source. Its 16 scalar states include two three-coordinate dynamic state sets selected from the four quaternion coordinates of bodyCylinder1.body and bodyCylinder3.body. Rumoca activates quaternion equations for those same two bodies; incorrect orientation-root activation does not explain this failure. The reference’s revolute and Rotor1D angular velocities both start at 10 and agree to about 4e-9 at the five-second endpoint.

At the final unreduced frontier, checked DAE function 30 is Frames.angularVelocity2. Its sole output assignment reads parameter ordinal 1, the whole three-element R.w input after record-argument lowering. It has no other statements or external implementation. Five body/rotor velocity equations call it, but equality closure reports no state anchor for their frame angular-velocity connectors. The missing fact belongs to structural equality analysis under SPEC_0007 / STRUCT-T03–04 and MLS FUNC-002–005, not to simulation or a model-specific exception.

The focused regression first fails because the checked forwarding function does not anchor its second vector argument to the state. Nonforwarding and asserting functions remain rejected by this proof. A separate four-vector Modelica fixture reproduces the structural failure at 9 of 12 matched rows: x = forward(a); a = b; y = forward(b); der(x) + der(y) = -2*x. Its solution is independently known as x = y = a = b = {1,2,3}*exp(-time). OpenModelica with the same Inline=true annotation used by MSL retains three scalar states and matches all twelve channels within 6e-9. Without that annotation, this OpenModelica build fails its array-equation code generation; that failed attempt is not counted as reference validation.

Equality closure now consumes the existing checked single-assignment forwarding proof and reads the returned caller argument with its original sign and shape. Original function and equation ownership is retained. All 146 structural unit tests pass. The ordinary originating-model attempt in target/msl/multibody-forwarding-gyroscopic exceeds the unchanged ten-second Solve phase budget and remains a failure. A separate diagnostic replay shows the final unmatched set shrink from 15 equations and unknowns to six of each: all nine unmatched angular-velocity connection rows disappear. The diagnostic replay is root-cause evidence, not a successful timed gate.

Once the forwarding fixture matches, state demotion leaves a repeated derivative in der(x) + der(y). Scalar Solve lowering previously accepted only an isolated derivative or one scaled derivative product, so the end-to-end regression then failed at that later boundary. It now derives the coefficient and offset symbolically for a proved scalar affine residual. The checked plan retains source parameter loads and emits ordinary arithmetic. It does not recover coefficients by subtracting two residual evaluations, which can lose them beside a large offset. Repeated-derivative tests use an offset of 1e20, change the runtime parameter, and require rejection of zero coefficients and nonlinear derivative products. All 116 Solve tests pass. The original four-vector fixture now agrees with the analytic trajectory on both BDF and RK, including all twelve initial values.

GyroscopicEffects: indexed matrix functions

With forwarding equalities present, the first refusal at both remaining orientation connection families is expression 1077, PromotedCat1, in checked function 33, Frames.from_Q. This function constructs its matrix from indexed quaternion products. Its single result assignment owns expression 1078. The diagnostic reports are .git/multibody-campaign/gyroscopic-orientation-proof-refusals.log and gyroscopic-forwarding-step-26-residue-12.dae.json in the same directory.

Structural differentiation now preserves the checked maps of linear tensor builtins and invariant index projections. Their original shapes and zero operands remain explicit; changing runtime indices acquire no such proof. Singleton equality substitution constructs the declared aggregate shape before a later projection, including when its anchor is scalar. This is the SPEC_0032 compact tensor rule applied inside the SPEC_0007 structural owner.

The matrix regression uses y = [x[2]^2, 7; x[1]^2, 11], der(x) = -x, and a = der(y), with the matrix expression owned by a Modelica function. A replay through the pre-map structural implementation reproduces 6 of 10 matched equations. Both solvers now match all ten analytic channels, including the constant column and its zero derivative. OpenModelica agrees within 4.2e-8. The original scalar-anchor projection regression also verifies successful reconstruction with the declared array shape. All 433 core tests pass with the final function-size refactor, as do all 116 Solve and 146 structural tests.

An earlier reduced probe also used vector() inside its function; that separately reaches the existing unsupported typed pure-call builtin boundary. The source is retained as .git/multibody-campaign/indexed-vector-unsupported.mo. The final matrix fixture follows the actual from_Q matrix-constructor form; support for that separate pure-call vector() boundary is not claimed. The new diagnostic replay in .git/multibody-campaign/gyroscopic-tensor-map-structural-report-2.log now clears the orientation connection frontier. It still stops with six unmatched equations and unknowns, now at the two three-dimensional frame_b.r_0 = frame_a.r_0 + Frames.resolve1(frame_a.R, r) families. The final holonomic candidate raises the unmatched count and is not accepted. This is the next proof boundary; the model still has no Rumoca trace. Focused all-target/all-feature structural and Solve Clippy passes after extracting the proof-walk operations into bounded functions. The fixed 20-model canary in target/msl/multibody-gyroscopic-focused-canary preserves all 20 phase, simulation, and agreement-band outcomes relative to multibody-initial-tensor-canary: nine comparisons are high, all 175 initial channels are high, and no comparisons are missing, skipped, or deviating. The artifact-bound delta is .git/multibody-campaign/gyroscopic-focused-canary-delta.json. Core Clippy also passes. None of these focused results establishes a new MultiBody or full-cohort parity count.

GyroscopicEffects: supplied kinematic derivatives (in progress)

The final tensor-map replay is retained in .git/multibody-campaign/gyroscopic-tensor-step-32-residue-12.dae.json. The first differentiation refusal in position rows 1400 and 1591 is a derivative coordinate: expressions 8454 and 8806 read der(Q) of bodyCylinder1.body and bodyCylinder3.body. Their quaternion derivatives are constrained implicitly by the angular-velocity equations. This rules out the earlier tentative trigonometric-builtin hypothesis at this frontier. The focused walk is gyroscopic-position-proof-refusals.log in that directory.

The MSL source for Frames.resolve1 supplies derivative(noDerivative=R) = Internal.resolve1_der and requests inlining after index reduction. Its supplied derivative uses resolve1(R, v2_der + cross(R.w, v2)), retaining the angular-velocity relationship instead of requiring a second quaternion derivative. The MLS §12.7.1 contract distinguishes omitting a tangent under documented argument assumptions from proving an argument constant. The proposed construction contract is in SPEC_0036 / SPEC_0043 §10, under Function Derivative Ownership.

The five-variable regression declares 2*q*der(q)=v, der(v)=-1, y=position(q,v), velocity=der(y), and acceleration=der(velocity). The position function returns q*q; its supplied derivative returns v under the stated kinematic relationship. With q(0)=v(0)=1, the exact position is 1+t-t*t/2. Rumoca fails at four of five matched equations in annotated-constraint-red.log. OpenModelica succeeds and matches all five channels within 8e-9, including initial values; the CSV and analytic receipt are under annotated-constraint-omc/. This records the original failure; the implemented DAE and structural derivative path now passes the regression.

Two earlier defects are reproduced and repaired: extraction discarded the actual parsed assignment form of restricted derivative annotations, and Resolve left their function and input references without declaration IDs. The red receipts are derivative-metadata-red.log and derivative-resolution-red.log. Resolve’s 152 tests pass after attaching identities in lexical scope. Flat metadata now retains exact callable references and ordered input roles, collects derivative dependencies, and expands each excluded record’s input role across exactly that record’s fields. The record-input integration test passes in derivative-flat-records-final.log. All 630 flattening unit tests and focused all-target/all-feature core, Resolve, and Flatten Clippy pass with the final metadata parser and role mapping. The preceding canary predates this work.

DAE schema 35 adds a checked differential owner. It requires complete, pure functions with matching original input names, types, and shapes, followed by precisely the required tangents. Record tangents preserve compact extents and recursively omit fields without Real values. The read-only view derives tangent input ordinals and the filtered result mapping; JSON and binary replay repeat the constructor checks. Structural reconstruction reattaches the links using its issued function identities. The preceding schema-34 implementation passed all 180 DAE and 146 structural unit tests in derivative-dae-structural-1.log. These include five derivative-owner tests for priority, omitted tangents, record types, malformed wire, and purity; the vector-state regression also verifies preservation during state demotion. All-target/all-feature DAE and structural Clippy passes in derivative-dae-structural-clippy-1.log; formatting and git diff --check pass.

Higher-order links now require a constructor-issued predecessor, retain source priority, and append tangents only for the preceding derivative group. Common zeroDerivative restrictions remain consistent. The six focused owner tests pass in derivative-chain-owner-1.log, including rejection of treating the second-order signature as an ordinary first derivative and rejection of a cyclic predecessor in wire data. Earlier schema-34 validation above predates this extension.

Shape discovery follows exact annotated callable identities separately from executable call dependencies. This allows a supplied derivative to call its primal without creating a false recursive component. Input value relevance also propagates through annotation dependencies. Structural first-derivative selection checks zeroDerivative applicability, retains omitted original arguments, and emits the selected checked function call before primal-body differentiation. Higher-order structural selection/emission is still pending; the checked chain alone is not an execution claim.

The restored first derivative closes the reproducer’s four-of-five structural failure. Its next rejection, recorded in derivative-integrated-core-3.log, was Solve’s requirement that the coefficient in 2*q*der(q)=v be compile-time numeric. Under MLS B.1 and SPEC_0007’s Solve lowering contract, scalar affinity does not require a constant coefficient. Scaled and repeated affine forms now retain state-dependent coefficients and evaluate the quotient on its finite, nonzero domain. A non-finite coefficient cannot yield a plausible zero result: the generated arithmetic reaches the runtime’s existing non-finite rejection. Declared zero/non-finite constant coefficients remain rejected at compilation.

derivative-integrated-core-4.log passes three tests: all five analytic channels with BDF and RK, refusal to apply zeroDerivative to the varying argument, and failure at the singular initial point q=0,v=1. Broad affected suites and a fresh GyroscopicEffects run were pending at that checkpoint.

derivative-focused-core-1.log now passes all 437 core tests. The fourth new regression verifies array specialization, dimensions used only by derivative functions, and predecessor identity for higher-order links; it does not claim higher-order structural execution. Affected library results are 181 DAE IR and 271 DAE lowering tests in derivative-focused-libraries-1.log, then 117 Solve and 146 structural tests in derivative-focused-libraries-2.log. The first library run exposed one outdated assertion banning every scalar Select; the repaired test instead proves the translation-time guard is absent from runtime dependencies and the coefficient remains live. Clippy found complexity violations in derivative replay, discovery, and proof walks; the affected operations are now separate helpers. The final four-crate Clippy run passes in derivative-focused-clippy-5.log. All 715 affected library tests pass after these refactors in derivative-focused-libraries-3.log, including the additional zero/non-finite scaled-coefficient cases. Core integration Clippy also passes in derivative-core-clippy-1.log.

The normal fresh-source run in target/msl/multibody-annotated-gyroscopic compiles GyroscopicEffects to a balanced 2,276-equation DAE in 3.32 seconds, then times out in Solve under the normal 10-second phase budget. It produces no trace, and the gate correctly reports parity unmeasured. Its log is derivative-origin-gyroscopic-1.log; this is a failure, not a coverage gain. The fixed 20-model run in target/msl/multibody-annotated-derivative-canary preserves every phase, simulation, and agreement-band outcome relative to multibody-gyroscopic-focused-canary. All nine comparisons are high and all 175 initial channels are high, with zero missing, skipped, excluded, or deviating comparisons. The artifact-bound receipt is .git/multibody-campaign/derivative-canary-delta.json, with working-tree digest d4a20c205dc516ba64a8c590aa2d6f40729a2d9dfca9c5cbfa5893e788f9f4d6. No new MultiBody or full-cohort coverage is claimed. verify quick, verify full, and the next full 566-model milestone remain pending.

A separate compile-only diagnostic exports fresh schema-35 source artifacts to .git/multibody-campaign/derivative-origin-diagnostic/. The DAE owns checked links for Frames.resolve1, resolve2, and resolveRelative, including the correct omitted tangents after record decomposition. The independently rebuilt structural inspector completes under its diagnostic 60-second cap; its report is derivative-origin-structural-report-1.log. The last exploratory reduction round has three unmatched equations, f_x[2138..2140] (Real T), and three unmatched unknowns: der(fixedRotation1.frame_b.R.w[2]), der(revolute.R_rel.w[3]), and bodyCylinder2.body.a_0[3]. It discovers no further direct or holonomic candidates. The position-equation frontier has advanced, but the failed reduction returns its original singular-system error; this diagnostic is neither a successful prepared system nor a simulation. The recorded normal-budget timeout above remains unchanged. The next focused step is to retain the final rejected candidate and inspect these three rows.

GyroscopicEffects: bilinear manifold values

The final exploratory candidate is retained as .git/multibody-campaign/derivative-final-candidate.dae.json, with its exact manifold and matching failure in the adjacent metadata file. The scratch pass copies the current structural implementation and adds diagnostics only; none of that instrumentation enters the production compiler. In the saved candidate, row 2138 has residual 8241, the connection between revolute.frame_b.R.T and fixedTranslation.frame_a.R.T.

The first refusal is expression 898, outerProduct(e,e) in Frames.planarRotation. Differentiation of the root succeeds, but retained value preflight rejects that builtin before it can reconstruct the record’s matrix field. The exact chain is in derivative-final-proof-refusals-3.log. The existing value emitter already supports the builtin; the preflight’s separate list omits it. This identifies the structural preflight as the first divergent layer and does not establish a state-selection or higher-order-derivative defect at this frontier.

MLS §10.3.5, catalog rows ARR-037/ARR-042, define the cross and outer-product operations. SPEC_0007’s structural contract requires retained manifold values to preserve those checked operations; SPEC_0029 requires one semantic helper owner. builtin_profiles now supplies the same value profile to preflight and reconstruction. DAE arity and shape checks, state-only operand proofs, and runtime-index refusals remain requirements of the path.

The reduced BilinearConstraint model returns a record containing outerProduct(x,v) + identity(3) and cross(x,v), constrains matrix/vector states to those fields, and observes their derivatives. Before the fix, bilinear-constraint-red.log fails at 27 of 39 matched equations. OpenModelica compiles and simulates that exact source, including its state-selection attributes, without warnings. All 39 channels match the independent exponential solution within 6.6e-8, including initial values; the source, CSV, log, and comparison receipt are in bilinear-constraint-omc/.

Restoring the value profile exposes a second reconstruction defect. The same model passes with Real literals for the function’s vector argument but fails with Integer literals, although the declared input is Real[3] in both cases. Both retained-manifold and direct-state substitution lose the declared Real type when replacing a formal parameter with the actual Integer expression. bilinear-constraint-types-3.log records those two failures and the successful Real-literal control. This rules out identity(3) and record shape as the source of the new type mismatch.

MLS §10.6.13 (ARR-009) requires Integer-to-Real conversion in Real contexts, including arrays. Structural substitution now preserves that source type in both value-rebuilding paths. The checked mixed-numeric identity 1.0 * value derives the Real result with the same compact extents; it introduces no scalar expansion or new IR vocabulary and does not relax builtin constructor checks. The existing initialization lowering uses the corresponding scalar identity value + 0.0 for the same language conversion.

All three reduced regressions now pass with BDF and RK, checking all 39 channels against the analytic solution, including initial conditions (bilinear-constraint-green-2.log). The broader run passes all 440 core, 146 structural, and 117 Solve tests (bilinear-focused-core-1.log and bilinear-focused-libraries-1.log). Structural Clippy passes with all targets and features (bilinear-focused-clippy-1.log); core Clippy also passes (bilinear-core-clippy-1.log).

The rebuilt production structural inspector now sorts the saved candidate after two direct substitutions (bilinear-final-structural-report-1.log). That snapshot omits the accumulated external manifold list, so this is a focused before/after proof, not a claim about complete initialization. The normal originating-model run supplies the stronger check: target/msl/multibody-bilinear-gyroscopic completes Solve construction in 8.715 seconds and reports ic_ok, then hits the unchanged 12-second simulation budget. Its gate fails with parity unmeasured: no Rumoca trace was produced. This advances the failure from Solve to simulation; it is not a model pass.

The fixed 20-model canary, target/msl/multibody-bilinear-canary, preserves all 20 phase, simulation, and agreement-band outcomes against multibody-annotated-derivative-canary. All nine compared models remain high, with 175 high initial-condition channels and zero missing, skipped, excluded, non-identifiable, or deviating comparisons. The durable receipt is .git/multibody-campaign/bilinear-canary-delta.json, binding the artifacts to worktree digest 1029dc260202fd272d4b3a4586bef573db23b0dc9de57769c9d55cbdb5465b32. These are focused results; no new full-cohort or MultiBody coverage number is claimed. The next investigation follows this same model into its first simulation steps.

The independent session probe (bilinear-runtime-probe-1.log) uses the original source DAE, the model’s 1e-8 relative and absolute tolerances, and a 12-second solver deadline. It records observations at 0, 0.01, and 0.02 seconds without changing the production compiler or the failed gate result. Integration advances in 38 and then 16 additional steps, with no root hits. All 967 exact name matches against the successful OpenModelica CSV in multibody-regression-restored-full agree within 2.5e-13 at initialization, 7.3e-8 at 0.01 seconds, and 4.4e-7 at 0.02 seconds. The other 1,309 Rumoca observation names have no exact CSV-name match in this diagnostic; this is neither full observable coverage nor a comparator agreement-band claim. Samples and the explicit comparison inventory are retained in .git/multibody-campaign/bilinear-runtime-samples/.

Profiling records 257,047 scalar-row evaluations through the second output, including 214,061 target-assignment evaluations. The emitted Solve model has 18 scalar states. Its derivative refresh contains 548 exact assignment rows but represents them through 479 projection stages: 477 singleton blocks and two coupled blocks of 79 and 72 equations. The algebraic refresh similarly contains 2,002 singleton projection stages and those two coupled blocks. The global causal-order certificate is false for both plans. In refresh_plan/schedule.rs, that global condition suppresses every exact assignment run, even around the coupled blocks. This is the next performance hypothesis, not yet a verified fix. A correction must prove dependency order for each local assignment run, retain the coupled solves, and preserve the existing rejection of uncertified seed order; simply removing the guard would discard the protection against stale algebraic values.

The six-variable CoupledRefresh regression confirms the scheduling defect: only a = sin(x) + b and b = 0.25*a are coupled, but all five algebraic variables enter projection stages. coupled-refresh-red-1.log records that failure while both solvers already satisfy the six-channel analytic solution. The proposed correction follows SPEC_0007 / SOLVE-C56 and SPEC_0029: construct local assignment stages from the existing exact-assignment dependency query, admit a row only after its dependencies in the projection inventory have settled, and retain the original BLT order where no global ordering proof exists. A coupled block remains a projection barrier. The existing checked Solve-owner stage-coverage and dependency checks still validate the result; runtime scheduling and tolerance rules do not change.

The local refresh schedule now batches assignments whose dependencies have been settled before each coupled solve. Validation: 417 IR/evaluator tests and 442 core tests passed, as did both focused Clippy checks. The six-channel analytic fixture agreed with OpenModelica to 2.733e-8. The fixed 20-model canary target/msl/multibody-local-refresh-canary preserved every phase, simulation, and comparison outcome against multibody-bilinear-canary (nine compared, all strict-high; zero missing, skipped, excluded, or deviating models; all 175 initial channels high). The worktree digest was 2670cb6c872308b21c3dcb66d5841287fd75977adf68a55434f189a42ac530f2. This is focused regression evidence, not a new cohort coverage claim.

The ordinary GyroscopicEffects run at target/msl/multibody-local-refresh-gyroscopic still timed out in Sim after successful Solve construction (8.840 seconds), backend construction (7.650 seconds), and initialization. Its comparator compared zero models, so parity remains unmeasured. A diagnostic replay showed fewer emitted stages but the same evaluation counts: the runtime correctly declined a schedule whose structural dependencies could invalidate earlier equations.

Tracing the first such dependency identified Solve projection block 703, bodyCylinder1.body.frame_a.R.w[1], as apparently invalidating rotation-matrix rows 1124–1132 in earlier blocks 600–608. The checked call owner and MSL Frames/from_Q.mo both show the matrix depends only on Q; w is forwarded to the separate angular-velocity output. StructuralPattern previously united every call input into every output, losing that separation. The reduced record_output_dependencies test reproduced the same false edge before the fix. Governing contracts are MLS §12.3 / FUNC-005 and SOLVE-C17/C51/C56; SPEC_0039’s dependency derivation describes argument substitution into a checked function summary.

Solve now derives a compact summary per typed output leaf from each issued body, substitutes nested summaries, and derives directional summaries from the same checked tangent body. Aggregates remain compact; within an aggregate the summary is conservative. Call-site claims are checked against these derived summaries during model replay (Solve schema 63). Runtime admission guards have not been relaxed. Validation of this dependency change is recorded below.

The reduced fixture also exposed an OpenModelica code-generation defect when its function lacked an inline annotation: the generated nonlinear residual copied four matrix entries to res+0 and two vector entries to res+1, overwriting matrix residuals and leaving two coordinates unwritten. Simulation reported success, but two trace channels were invalid. That run is retained as failed oracle evidence in record-dependency-omc. The regression now uses Inline=true, matching MSL from_Q, without changing its equations; the exact nine-channel source in record-dependency-inline-omc agrees with its analytic solution at all 12 output rows to 4.441e-16. Both oracle artifacts are under .git/multibody-campaign/; this is diagnostic evidence, not an MSL exclusion.

The dependency change passes 255 IR and 167 evaluator tests, all 445 core tests, and both focused Clippy checks. The ordinary originating-model run, target/msl/multibody-record-dependency-gyroscopic, constructs Solve in 9.345 seconds and its backend in 7.972 seconds, initializes successfully, and still hits the 12-second Sim limit. No trace comparison is available from that run. The fixed canary target/msl/multibody-record-dependency-canary preserves all 20 outcomes against multibody-local-refresh-canary: nine compared, all strict-high; zero missing, skipped, excluded, or deviating models; all 175 initial channels high. Its recorded worktree digest is a6e671e284a11effa5da2f0933d9f1511c7454a6fd45e0d8d3c95359e4127146.

The full-model diagnostic confirms the original angular-velocity-to-matrix edge disappeared. Remaining reverse-invalidation flags fell from 21 to 14 for algebraic refresh and 15 to 10 for derivative refresh. Evaluation counts remain 257047 scalar rows and 214061 target assignments through time 0.02; all 2276 observables at times 0, 0.01, and 0.02 are unchanged from the pre-fix diagnostic. These are three samples, not a full trace or a new model pass.

The next false edge is at block 704, updating velocity component 2 after component 1 in block 703. Logical implicit row 1388 maps through ScalarPrograms.output_indices to stored program 424. That program calls owner 141, the three-operation body of Frames.angularVelocity2: load the rotation matrix, load angular velocity, and return angular velocity. The new summary correctly omits the matrix but still unites the three vector components. The next construction obligation is to retain compact component mappings through such typed copies and their directional/nested calls. Incomplete projection seed coverage remains a separate admission condition. Neither guard can be removed merely to enable the optimized path. Full 42-example, 566-model, and quick/full verification remain pending while this focused runtime frontier is unresolved; the complete MultiBody goal and PR gate are unchanged.

The component-copy regression now reproduces that second false edge directly: each returned vector coordinate previously depended on all three input coordinates. Solve schema 64 retains compact affine coordinate relations in the owner-derived output summaries. Typed copies, pointwise operations, transpose, static projections, and matrix products propagate their checked coordinate relations; nested calls compose them, including independent contraction axes. The same derivation applies to the checked directional body. Other operations retain explicit conservative input dependencies. Coordinate enumeration occurs only in the scalar structural-pattern view. The governing contracts remain SOLVE-C17/C51/C56 and SPEC_0032 §6, with MLS §10.6.1 / ARR-027 for array assignment and §12.3 / FUNC-005 for pure calls.

Ten dependency tests pass, including all four matrix-product rank combinations, nested contractions, transpose/slice composition, primal/tangent separation, forged wire claims, and metadata independent of a million-element extent. All 260 Solve IR and 167 evaluator tests passed before the mechanical Clippy helper extraction; all 447 core tests and both focused Clippy checks pass after it. The new reduced source solves through angularVelocity2 in the same direction as GyroscopicEffects and checks all eight observables with both solvers and both execution policies. Its exact OpenModelica source in .git/multibody-campaign/component-copy-omc has 12 output rows, includes the initial point, and agrees with the analytic solution to 1.3573e-10 absolute error.

The current full-model diagnostic removes all remaining reverse-invalidation flags: algebraic 14 to zero, derivative 10 to zero. Scalar-row evaluations through time 0.02 fall from 257047 to 229999, and target assignments from 214061 to 187349, with the same 54 integration steps. All 2276 observables are identical at initialization and differ by at most 1.095e-12 at the two later sample times. These remain three diagnostic samples, not a full trace. The normal run target/msl/multibody-component-dependency-gyroscopic constructs Solve in 9.926 seconds and its backend in 8.043 seconds, initializes, and still times out after the standard 12-second Sim budget. It compares zero models; parity remains unmeasured. The remaining staged-refresh admission failure is projection seed coverage. The first projection block, 703, has target Y867 and no seed row; coupled blocks also retain incomplete seed coverage. This needs a separate correctness argument and regression before changing admission.

The fixed canary target/msl/multibody-component-dependency-canary preserves all 20 phase, simulation, and band outcomes against multibody-record-dependency-canary. Nine models are compared, all strict-high, with zero missing, skipped, excluded, nonidentifiable, or deviating models; all 175 initial channels are high. The receipt .git/multibody-campaign/component-dependency-canary-delta.json binds worktree 56d06da8111e4d948cd1620c007aa88760998934591ef0e9ddc5b34b76a2ed8e. This is Tier 1 regression evidence, with no new model pass or cohort claim. Full 42-example, 566-model, and quick/full verification remain outstanding while the originating model still fails; no new commit or PR has been made.

The first missing seed is now reproduced and repaired at the assignment certificate owner. GyroscopicEffects row 1388 declares Y867 as its target, but Frames.angularVelocity2 returns that coordinate through a typed call, which the isolator previously could not follow. The reduced copied-record fixture failed because no exact refresh assignment existed for R.w[1]. Solve schema 65 derives a separate exact value-projection certificate from checked copies, finite constants, static projections, fills, transposes, and nested certified calls. Arithmetic, dynamic indexing, control flow, and assertion-bearing bodies remain outside this proof. Dependency information alone never establishes value equality. Assignment isolation consumes the certificate while retaining the call’s evaluation prefix; runtime admission checks are unchanged. Wire replay rederives the certificate and also binds the paired directional call to its exact owner, closing a separately reproduced owner-identity forgery.

Validation passes 268 Solve IR, 167 evaluator, and 449 core tests, plus both focused Clippy checks. The reduced source remains byte-identical to the eight-channel OpenModelica fixture above and passes both solvers with both execution policies. Positive and negative tests cover nested coordinate composition, compact million-element metadata, evaluation-prefix retention, nonlinear rejection, and forged primal/directional wire claims.

The full-model diagnostic restores six exact velocity seeds. Derivative refresh blocks with incomplete seed coverage fall from eight to two; those remaining coupled blocks contain 79 and 72 unknowns. Algebraic refresh still has 51 blocks with incomplete coverage. Scalar-row evaluations through time 0.02 fall slightly, from 229999 to 228861; target assignments remain 187349. Across all 2276 observables, initialization is unchanged and the two later samples differ by at most 1.140e-11. These are diagnostic samples, not a complete trace. The normal originating run target/msl/multibody-forwarded-value-gyroscopic constructs Solve in 9.862 seconds and its backend in 8.572 seconds, initializes successfully, and still hits the standard 12-second Sim limit. Zero models are compared, so parity remains unmeasured and no additional model pass is claimed.

The fixed canary target/msl/multibody-forwarded-value-canary preserves all 20 phase, simulation, and band outcomes against multibody-component-dependency-canary. All nine compared traces remain strict-high, with zero missing, skipped, excluded, nonidentifiable, or deviating models and all 175 initial channels high. The receipt .git/multibody-campaign/forwarded-value-canary-delta.json binds worktree e1276912dd733b7489f2faf313b5307c9147c9689a4428804dd57f918f687504. This closes Tier 1 validation of the value-projection change. The remaining coupled-block admission failure needs its own source reproduction and correctness argument; complete 42-example/566-model and quick/full verification are still pending. No new commit or PR has been made.

The next reduced defect is block-relative affinity. The exact schema65 GyroscopicEffects artifact in .git/multibody-campaign/forwarded-value-runtime-samples/solve.json (SHA256 b84179b46a0324e4ae8bb256b9c3a1e2b9306e539204ee41dd1692b9a2cfbe18) contains coupled blocks 1794 and 1855, with 79 and 72 unknowns. At fixed outside-block values, every selected residual is affine in its own block’s unknowns. The saved OpenModelica regular systems 1514 and 1760 are also classified linear, but their different state and alias selection does not establish a one-to-one block mapping. Initial quaternion equations remain nonlinear. The diagnostic receipts are coupled-block-affinity.json and coupled-block-omc-systems.json in the campaign directory.

The actual runtime predicate previously required a parameter-static gradient with respect to all solver variables. A two-variable regression with an earlier block supplying a coefficient rejects that unnecessarily strong condition. Solve schema66 derives compact typed-call input-interaction summaries and checks residual degree against the exact canonical block inventory. Wire replay rederives the facts and rejects altered call claims or mismatched canonical rows and unknowns. The runtime uses the issued fact to remove the block’s Newton step cap; it proves neither nonsingularity nor parameter-static gradients, and does not change staged-refresh seed admission. Tensor ranges stay compact. Unsupported compute forms remain unproved.

Review reproduced an output-identity error in the new checker before acceptance: two compute nodes each using local output zero swapped the affine/nonlinear classification. block-affinity-output-identity-red-1.log retains that failure. The corrected checker uses the existing compute-block output mapping and refuses duplicate output ownership. All 281 Solve IR, 167 evaluator, and 418 solver tests now pass in block-affinity-libraries-final.log; all 452 core tests pass in block-affinity-core-full-1.log. The reduced tensor fixture checks all nine analytic observables with both BDF/RK and Auto/Interpreter. Its exact source SHA256 is 93928e0f5f69b4f37e9859c1915c79298501e444420a683f63769003f3c65d5f. Default OpenModelica tearing refuses the non-inlined call while generating linear-system Jacobians; that failure remains in block-affinity-omc-1.log. The identical source succeeds with --tearingMethod=noTearing (the pinned OMC version’s deprecated alias for minimalTearing): all nine channels and 12 rows, including initialization, agree with the analytic solution to 2.142e-10 absolute error. The alternate-oracle receipt is block-affinity-no-tearing-omc/analytic-comparison.json; this diagnostic changes neither MSL references nor gate settings. Both final focused Clippy checks pass, and temporary profiling tests have been removed.

The production diagnostic certifies both actual coupled blocks affine and retains 18 states. Its short run produces bitwise-identical values for all 2276 observables at times 0, 0.01, and 0.02, with unchanged work counts: 228861 scalar-row evaluations, 187349 target assignments, and 54 solver steps. This is not a complete trace and shows no speed gain. The normal run target/msl/multibody-block-affinity-gyroscopic compiles in 3.388 seconds but exceeds the 10-second Solve-construction budget; the parent observes the failure after 13.644 seconds. This is an earlier failure than the preceding Sim timeout, with no initialization or trace result. It remains an unresolved performance regression; the run compares zero models and parity is unmeasured.

The fixed canary target/msl/multibody-block-affinity-canary preserves all 20 phase, simulation, and band outcomes against multibody-forwarded-value-canary. All nine compared models remain strict-high, all 175 initial channels are high, and missing, skipped, excluded, nonidentifiable, and deviating counts are zero. The receipt block-affinity-canary-delta.json in the campaign directory binds worktree 1e78ae01f6a0f6e65b247bb3367da214ca7122a54d4809cba2292575e7fc4d79. This is Tier 1 regression evidence, with no coverage gain.

Isolated construction measurements rule out degree derivation as the main multi-second cost: all 2004 projection proofs take 3.4–3.7 milliseconds, and primal/directional interaction derivation over 1967 function owners takes 6.0–6.8 milliseconds. Replaying the complete checked function table takes 6.53–6.70 seconds in the test profile. These measurements do not include all interface-cloning costs in the degree-derivation totals. The before/after function-table wire is identical, and the 251 root owner references reach all 1967 owners through typed call edges. Unused-owner pruning is therefore not a remedy for this inventory. The next hypothesis is repeated construction and deep cloning of previous-owner interfaces, which needs direct measurement and a reduced regression before changes. Complete 42-example/566-model and quick/full verification remain outstanding; no new commit or PR has been made.

The interface-construction hypothesis is now measured. Rebuilding all previous owner interfaces for each of the 1967 owners takes 682.699 milliseconds for primal signatures and 3.687 seconds for directional interfaces alone (call-interface-prefix-profile-before.log). The borrowed-storage regression fails on the former interface() implementation because it copies the owner’s input array (call-interface-borrow-red-1.log).

Typed construction now uses a Copy borrowed SolvePureCallTableView over already-issued owners. Constant-time lookup retains the exact owner ID and selects primal or directional interfaces without allocating previous-owner arrays. Program construction, nested regions, wire replay, and all three summary derivations share that view. Future owner IDs and suffix/index rebinding are rejected. The finalized program and schema66 wire remain owned; this change neither skips wire checks nor merges function identities.

All 284 Solve IR, 167 evaluator, 418 solver, and 452 core tests pass. All nine Solve IR doctests pass, including an actual E0521 rejection of a register crossing construction scopes. Both affected-library and compiler Clippy pass; formatting and git diff --check pass.

The fresh diagnostic lowers and saves the model in 8.677 seconds, versus 10.204 seconds before this change, about 15% faster. The complete serialized Solve artifact is byte-identical, with SHA256 240ca32f5e37ec32f45f435594d0f8a994de904b1255ff8ebbb35f492ec9c8c4. The ordinary run target/msl/multibody-call-interface-gyroscopic now clears the unchanged construction budget: compilation takes 3.477 seconds, Solve construction 8.725 seconds, and backend preparation 7.685 seconds. Initialization succeeds, but simulation still exceeds its 12-second budget. This repairs the construction timeout, with no additional model pass: zero models are compared and parity remains unmeasured.

The fixed canary target/msl/multibody-call-interface-canary passes and preserves every phase, simulation, and band outcome of the preceding multibody-block-affinity-canary. All nine compared models remain strict-high; all 175 initial channels remain high. Missing, skipped, excluded, nonidentifiable, and deviating counts are zero. The campaign receipt call-interface-canary-delta.json binds worktree ecf962178d1838402b99034a846ab01ba621b84191d3f821180793700e4df5f1. This is Tier 1 regression evidence, not a cohort coverage gain.

The remaining investigation is runtime cost in repeated algebraic projection, including reduced Jacobian sweeps. Block affinity alone does not justify reusing finite-difference Jacobians or removing causal-seed admission; both proof boundaries remain intact. Complete 42-example/566-model and quick/full verification remain outstanding while the originating timeout is unresolved. No new commit or PR has been made.

Runtime profiling next isolated a separate allocation defect in rumoca-eval-solve::typed_program::InvocationScope. Every function invocation and Map/Fold domain point created and destroyed a result array sized for every function owner in the model, including leaf calls with no nested calls. Of 209 CPU samples in a final 1.05-second integration window, 27.27% were in that array’s initialization and 13.40% in its destruction. These are self-symbol samples, with incomplete stack unwinding, from a short diagnostic rather than a full-model benchmark. runtime-projection-perf-integration.txt and invocation-scope-triage.json retain the attribution and limitations.

The resource regression in invocation-scope-red-1.log invokes a leaf through the actual evaluator in a checked 1024-owner table: its correct output still retains 1024 unused result slots. Invocation storage now contains only executed calls, keyed by the same typed owner ID. The former owner bound remains an explicit check; duplicate insertion fails without replacing the issued result. The lookup is private and never determines output order. Conditional regions still share their enclosing invocation, while function and Map/Fold scopes remain fresh. No value survives its existing invocation lifetime. A leaf’s result lookup now has zero allocated capacity, and a high owner ID consumes only one entry when actually invoked. Unknown and duplicate owner regressions pass. This changes evaluator storage, with no IR, arithmetic, solver tolerance, or seed-admission changes, under SPEC_0029 sections 5/12 and SPEC_0043 sections 6/9’s existing execution, call-identity, and scope contracts.

All 169 evaluator and 418 solver tests pass in invocation-scope-libraries-1.log; all 452 core integration tests pass in invocation-scope-core-1.log. The all-target/all-feature evaluator/solver Clippy check, formatting, and git diff --check pass. The fresh diagnostic’s two integration intervals take 503.549 and 213.184 milliseconds, versus 910.202 and 380.146 milliseconds in the preceding profiled run, about 44% less time. Work remains exactly 228861 scalar-row evaluations, 187349 target assignments, and 54 solver steps. All 2276 observables at times 0, 0.01, and 0.02 remain bit-identical, including the serialized sample files (invocation-scope-runtime-comparison.json). A separate direct-name check against the saved OpenModelica CSV covers 967 observables at these times, with maximum absolute error below 4.391e-7; 1309 Rumoca channels are unmatched in this diagnostic without alias expansion. This incomplete three-sample check is not a trace-parity band or a model pass.

The normal target/msl/multibody-invocation-scope-gyroscopic run compiles in 3.252 seconds, constructs Solve in 8.648 seconds, and prepares the backend in 7.056 seconds. Initialization succeeds, but simulation still exceeds its unchanged 12-second budget. Zero models are compared and parity is unmeasured. The fixed target/msl/multibody-invocation-scope-canary passes with every one of its 20 phase, simulation, and band outcomes unchanged from multibody-call-interface-canary: nine compared models remain strict-high, all 175 initial channels remain high, and missing, skipped, excluded, nonidentifiable, and deviating counts remain zero. The Tier 1 delta receipt invocation-scope-canary-delta.json binds worktree c522ffb348ed1e921b780fe1ccc1bb362969741ed66025d5b865be3b3bb29a79. There is no additional model pass or cohort coverage claim. Full 42-example, 566-model, and quick/full verification remain outstanding.

The follow-up profile uses a realtime-clock window bounded by the first and last sample markers: 361 CPU samples over 0.724 seconds. The removed dense invocation-array symbols are absent; the largest remaining self symbol is typed EvalFrame::run at 15.51%, with typed arithmetic and value allocation also prominent (invocation-scope-perf-integration.txt). Source inspection shows that RefreshProjectionModel::eval_implicit_target_value always enters the interpreted target-assignment evaluator. The next investigation is that execution boundary and the existing compiler-issued exact assignment/native call products. This is a hypothesis for the next fix, with no change yet to projection scheduling, nonlinear seeds, or Jacobian policy.

The next reduced regression confirms a missing execution binding: RefreshProjectionModel::eval_implicit_target_value passes the model’s row context to the exact target-assignment evaluator, but that context supplied only the reference call table. Both scalar-call evaluators consequently interpreted typed functions even when the native backend already owned their compiled helpers. The fixture uses the real compiled table for f(x)=x*x+2 and the actual target-assignment evaluator. Before wiring the consumer, it returns 11 at x=3 but records zero native invocations where one is required (native-call-projection-red-2.log). The preceding red-1 log is only a fixture compile error, not regression evidence. This establishes execution dispatch as the responsible layer: checked call construction and the numerical result are already correct, and the compiled table is available in this fixture.

The evaluator context now accepts the model backend’s pure-call executor. Both checked and prepared scalar paths validate the same issued call site and marshal through the same typed input conversion as reference evaluation. The Cranelift adapter reuses the existing model-wide primal or directional helper and its checked ABI; it builds no per-target wrapper or duplicate IR. Outputs are invocation-local and are published only after successful execution. An attached executor’s error propagates with the source span instead of retrying reference evaluation. The interpreter policy still supplies no native executor. Governing rules are SPEC_0029 sections 5/12 and SPEC_0007 Stage 4, especially SOLVE-C33/C37/C45/C51/C56; this is execution binding, not a new Modelica transformation, scheduling rule, or result-cache lifetime.

All three new regressions pass (native-call-projection-regressions-2.log): the real compiled owner executes on repeated target assignments with changing inputs; primal/directional outputs agree with reference evaluation and the analytic derivative for changing inputs and seeds through both row entry points; and an injected execution failure publishes no partial output and retains its source span. The complete affected library run passes 169 evaluator, 66 Cranelift, 127 simulation, and 418 solver tests, and all 452 compiler core tests pass: 1232 in total. All-target/all-feature Clippy for the four affected libraries, formatting, and git diff --check pass. The library run also found an inspection test expecting the old fixed-start pendulum refusal. Its positive replacement checks the issued given-state ownership, inspection agreement, exact initial x=1, and the trajectory’s circle constraint. A separate contradictory fixed-start alias fixture retains the reduction, inspection, and diagnosis refusal checks. All five structural-report tests pass. The fresh short diagnostic takes 116.603 and 51.368 milliseconds for its two integration intervals, versus 503.549 and 213.184 milliseconds before binding the native calls: about 76–77% less time. All 2276 values at times 0, 0.01, and 0.02 remain bit-identical, including the complete sample JSON files (native-call-runtime-comparison.json). The solver still takes 54 steps and evaluates 187349 target assignments. These three samples are not a full trace or a parity band.

The normal target/msl/multibody-native-call-gyroscopic run compiles in 3.974 seconds, constructs Solve in 9.168 seconds, and prepares the backend in 6.173 seconds. Initialization succeeds, but simulation still reaches its unchanged 12-second timeout. The failure bucket is Timeout; the raw sim_solver_fail tag does not establish a new numerical failure. Zero models are compared, and parity remains unmeasured.

The fixed target/msl/multibody-native-call-canary passes with all 20 phase, simulation, and band outcomes unchanged from multibody-invocation-scope-canary. All nine compared models remain strict-high, all 175 initial channels remain high, and missing, skipped, excluded, nonidentifiable, and deviating counts are zero. native-call-canary-delta.json binds worktree 22295d5f1c68acb5c36c1f742917f53c1f44039663ef6340ffc954aaa73f7507. This is completed Tier 1 evidence for the execution binding, with no additional model pass or cohort coverage claim. Complete 42-example/566-model and quick/full verification remain outstanding; no new commit or PR is made.

Fresh 42-example checkpoint after native call binding

target/msl/multibody-native-call-examples completes the unchanged 42-example target list at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, with worktree digest c5ccf781508b2d716b052ea45022e60275984fdbc101d598ca94b3f3128ed7d2. All 12 completed models are compared and strict-high across 4837 channels, including initialization. Missing, skipped, excluded, nonidentifiable, and deviating counts are zero. FreeBody and ThreeSprings change from absent to high; all ten previously high examples remain high. This comparison spans the accumulated branch changes since multibody-record-field-examples, so the two gains cannot be attributed solely to native execution binding. These are measured example results, not a new golden-cone certification or a 566-model cohort claim.

The remaining failures stay in the target list:

The campaign receipts native-call-examples-receipt.json and native-call-examples-delta.json retain the exact failures and phase/band changes. This replaces the older 10-of-42 checkpoint for this example scope. The broader library/test inventory and complete 566-model and quick/full gates remain unfinished.

The subsequent longer GyroscopicEffects diagnostic advances to time 0.9079194955562074 after 392 solver steps before reporting the same 12-second timeout. A realtime-bounded profile contains 5032 self-symbol samples over 10.090 seconds: native scalar Jacobian rows account for 39.07% and native typed directional helpers for 32.91%. There are no lost samples. These are self-symbol classifications, not a call-tree attribution; repeated JIT symbol names are not semantic owner identities. The shorter observed advance interval versus the reported solver timeout also remains to be explained. native-call-full-perf-window.json records the window and qualifications. The next investigation is the actual Jacobian caller and selected-row/full-JVP dispatch, with a reduced reproduction before changing execution or reuse.

The next stack trace stops at the first native Jacobian evaluation after the private probe’s integration marker. The actual call chain is BDF state reset and Jacobian assembly → FMI directional derivative → projected state sensitivity → algebraic_plan_row_scales → algebraic_block_jacobian → the compiled full-model JVP (jvp-callstack-1.log). The colored branch of the shared block assembler always invoked the full JVP; unlike the ordinary branch, it never requested the available selected-row evaluator. This is runtime execution ownership under SPEC_0029 sections 5/12 and SOLVE-C17/C45/C56, not an IR sparsity or differentiation change. The existing coloring still proves that columns in a color affect disjoint rows.

The reduced colored_rows fixture has three local rows mapped to global rows 4/1/3, local columns mapped to Y[2]/Y[0]/Y[1], and two checked colors. The actual row-scale helper returns the correct scales but makes two whole-model JVP calls rather than selecting the four required row/color pairs. All four tests fail before the fix (colored-jvp-red-2.log); red-1 is only a fixture compile error. The colored branch now passes its active global rows to the same selected-row helper as ordinary projection. Local matrix placement, reverse-computed rows, coloring, parameter-zero seed lanes, and scaling are unchanged. An unavailable row retains one full evaluation for that color; an error propagates rather than selecting that fallback.

All four focused regressions, all 422 solver library tests, and all 452 core tests pass, along with all-target/all-feature solver Clippy, formatting, and git diff --check. All 2276 short-diagnostic values at 0/0.01/0.02 remain bit-identical (colored-jvp-runtime-comparison.json). The normal multibody-colored-jvp-gyroscopic run clears initialization, with Solve construction at 8.735 seconds and backend preparation at 4.512 seconds, but still reports a 12-second simulation timeout: zero compared, parity unmeasured. The fixed multibody-colored-jvp-canary passes with all 20 outcomes unchanged from multibody-native-call-canary: nine compared strict-high models, 175 high initial channels, and zero missing/skipped/excluded/nonidentifiable/deviating counts. The delta receipt binds worktree ba078c68234cc2490ad2541b1c68b3681b5ef0cdddb14020edfa475ff1f7832d.

The longer diagnostic now reaches time 4.48058911338337 after 1906 steps, versus 0.9079194955562074 after 392 steps before this change. It still times out under the unchanged budget. Its realtime window has 5801 self-symbol samples over 11.634 seconds, with zero samples in the formerly dominant native Jacobian-row functions (colored-jvp-full-perf-window.json). This completes Tier 1 evidence for selected colored JVP execution, without claiming a model pass. Inspection of the same sensitivity path also finds that it reconstructs each block’s Jacobian for final row scaling after having formed it at the identical immutable Y/P/time point for the sensitivity solve. The next reduced check tests retaining those row scales from the existing matrix within that one invocation; it does not propose cross-step caching.

The reduced two-row sensitivity case confirms the duplicate work: twelve selected JVP calls instead of eight (seed-scale-reuse-red-1.log). The primal Y/P slices and time remain fixed while only the tangent seed changes, so final row scaling can consume the same matrix used to solve each block. Review of the initial implementation caught a contract regression: its separate seed matrix builder did not validate the structure dimensions that the old final scaling pass checked. A mismatched 3x3 structure attached to a 2x2 block was incorrectly accepted (seed-scale-shape-red-1.log).

Sensitivity projection now uses the shared algebraic_block_jacobian and computes row scales before LU consumes that matrix. The duplicate per-column builder and its unused unit-seed scratch plumbing are removed. Shape checking, selected-row errors, singularity rejection, final scaled residual validation, and restoration of the caller’s unknown seeds on failure remain in place. This is runtime execution ownership under SPEC_0029 sections 5/12 and SPEC_0007 Stage 4, not a new compiler transformation or cross-call cache. The colored fixture also checks a nonzero projected direction through the nonidentity row/column layout while retaining the caller’s known directions. All 424 solver library tests and 452 compiler core tests pass (seed-scale-reuse-solver-3.log, seed-scale-reuse-core-1.log), along with all-target/all-feature solver Clippy, formatting, and git diff --check. The short diagnostic retains bit-identical values for all 2276 channels at 0/0.01/0.02 (seed-scale-reuse-runtime-comparison.json), with the same 54 solver steps. The normal multibody-seed-scale-gyroscopic run compiles in 3.268 seconds, constructs Solve in 8.612 seconds, and prepares the backend in 4.427 seconds. Initialization succeeds, but simulation still reports the unchanged 12-second timeout: zero compared and parity unmeasured.

The fixed multibody-seed-scale-canary passes with every phase, simulation, and band outcome unchanged from multibody-colored-jvp-canary. All nine compared models are strict-high, all 175 initial channels are high, and missing, skipped, excluded, nonidentifiable, and deviating counts are zero. seed-scale-canary-delta.json binds worktree 5cd1f9433dfd1ed236bf84fe2a4f8b89b897bc15ae69b2ce79cb5a0ff2f2435a. This completes Tier 1 validation for this runtime change without adding a full-model parity claim. The complete 42-example checkpoint remains twelve strict-high examples; complete 566-model and quick/full verification remain pending.

The subsequent longer diagnostic reaches time 4.900310731426584 after 2098 steps, versus 4.48058911338337 after 1906 steps before matrix/scale reuse. It still reports the same timeout. seed-scale-full-perf-window.json records 5871 self-symbol samples over 11.775 seconds. The profile includes 2.37% in certifies_exact_target_assignment_output; an actual debugger stack places that call in branch-continuity projection during a BDF derivative evaluation (assignment-certification-callstack-1.log). Inspection finds it rescanning the prepared block’s immutable operation list for non-causal operations on each query. This is the next reduced performance investigation, not an implemented optimization. The profile’s separate metadata-comparison cost does not authorize removing call-site ownership checks.

The reduced prepared-block fixture contains two independent assignment outputs and one seed-dependent residual. Its valid/invalid row, output, and target checks pass before the resource assertion fails: repeated queries over the original and cloned block perform 408 operation visits after preparation (prepared-causality-red-1.log). The predicate reads only immutable owned operations, so its result belongs with the evaluator’s existing prepared row metadata. The output/target shape checks remain specific to each request.

PreparedScalarProgramBlock::new now records the existing non-causal-operation predicate once per program. Its metadata allocation remains fallible, and cloning retains the same result for the cloned program. Prepared certification and isolator materialization consume that result. Standalone program helpers still check their raw input. The regression also checks materialization and proves that preparation performs the scan while subsequent queries and cloning perform none. The per-thread counter exists only in unit-test builds. This is evaluator preparation under SPEC_0029 sections 5/12 and SPEC_0007 Stage 4; it changes no accepted operations, runtime coordinates, call-site ownership, or arithmetic. All 170 evaluator, 424 solver, and 452 compiler core tests pass (prepared-causality-libraries-1.log, prepared-causality-core-1.log), along with all-target/all-feature evaluator/solver Clippy, formatting, and git diff --check. The MSL results follow below.

The short diagnostic retains all 2276 values bit-identically at 0/0.01/0.02. The fixed multibody-prepared-causality-canary passes with all 20 outcomes unchanged, nine compared strict-high models, 175 high initial channels, and zero missing/skipped/excluded/nonidentifiable/deviating counts. Its receipt binds worktree 130cbc109b21ef9d040a23b119ffb26ce31e606b4128885219de7bb8e0607399.

The complete multibody-prepared-causality-examples run has 13/42 strict-high models and 5754 high channels including initialization, with zero missing, skipped, excluded, nonidentifiable, or deviating channels/models. All twelve previous high examples remain high; Constraints.SphericalConstraint newly passes all 917 compared channels. Its simulation takes 7.484 seconds. These gains span the colored JVP, matrix/scale reuse, and prepared-causality changes since multibody-native-call-examples, not just the last edit. GyroscopicEffects still reaches the unchanged simulation timeout after successful initialization; the origin remains unmeasured. The 29 remaining example failures comprise 15 timeouts, six structural-analysis failures, six DAE-construction failures, one initialization runtime-contract failure, and one instantiation failure. The example receipt and delta preserve each result.

An independent source-grounded review of SphericalConstraint checks the two equivalent spring/body assemblies over all 501 samples from 0 to 10 seconds. The maximum locked-position residual is 2.17e-16 m; the relative-position sensor residual is 9.73e-15 m; constraint torque and power are zero. Rotation orthogonality residuals are at most 2.10e-8 for the joint body and 6.66e-16 for the constrained body. Computing mechanical energy from the source-defined mass, inertia, center of mass, spring, and gravity gives maximum drifts of 4.63e-5 and 5.91e-5 J from the common initial 1.91568 J. The largest difference between corresponding body position/orientation/angular-velocity channels is 8.37e-5. spherical-physics-review.json records raw maxima and source/trace hashes; these diagnostics are additional physical evidence, not a new tolerance policy or complete golden IR-cone certification.

The subsequent complete 566-model run multibody-runtime-prepared-full, at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311 and the same worktree digest, fails the soundness gate after the comparator executes. It measures 142 models: 141 strict-high, one near, zero missing, and 18 skipped under existing reviewed policy exclusions. All 138 previously high models remain high. FreeBody, ThreeSprings, and SphericalConstraint account for the three high gains. The additionally completed single-phase RLV_Characteristic activates an already tracked exclusion; no exclusion policy was changed or baseline promoted. All 14550 compared initial channels are high.

The newly completing Electrical.PowerConverters.Examples.ACDC.RectifierBridge2mPulse.ThyristorBridge2mPulse_RLV has 409 high and 20 deviating channels, 17 severe, despite its aggregate near band. This is one actionable refinement counterexample and takes priority over further MultiBody performance/capability work. Its largest errors are aliases of neutral-point voltage: Rumoca’s error reaches roughly 0.003 V while the OMC reference range is about 1.5e-9 V. The source connects the star point through a 1e6-ohm resistance to ground. This suggests sensitivity to small current-balance errors, but the responsible compiler/runtime layer is not yet proven. No noise exception, tolerance change, or closure is claimed. runtime-prepared-full-delta.json retains the complete phase/band delta; the next task is a reduced reproduction and root-cause proof for this model.

A separate diagnostic worker reproduces the neutral-voltage error and saves Flat, DAE, structural DAE, and Solve artifacts without changing the model or solver settings. At t=0.0594 the neutral voltage is 0.002971540144 V and the 1e6-ohm resistor current is 2.971540144e-9 A, retaining the resistor’s exact constitutive relation. The first sample above 1e-6 V occurs at t=0.002. The neutral voltage/current belong to Solve projection block 50, a 110-unknown coupled block with ten tear variables; the reported ground current is a later singleton assignment. This narrows the next investigation to how the coupled projection converges and how its settled values reach observation. It does not yet distinguish residual-only convergence, ill-conditioned numerical elimination, or an incorrect producer from one another.

Broad verification checkpoint

The requested cargo xtask verify quick run at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311 fails four stages: MSL parity, the corpus pin, architecture, and workspace tests. Workspace-wide all-feature Clippy and formatting pass. The MSL results are preserved in target/msl/multibody-verify-quick-checkpoint-1, with worktree digest 1acce4d1c34423640ac465849a4300cf34ac817dcc2a1b5c2c89c25f6e4c1e18. The complete 566-model comparison retains all 141 previously strict-high models: 142 compared, 141 strict-high, the same one rectifier counterexample, 17 existing policy exclusions, zero missing, and zero nonidentifiable. No model changes comparison band. The already excluded single-phase ThyristorBridge2Pulse_RLV_Characteristic times out after completing in the preceding run; that execution loss remains recorded separately.

The workspace runner initially stops after 372 tests on the architecture failures. A subsequent run with --no-fail-fast --test-threads 4 executes all 7294 tests: 7290 pass, four fail, and none are skipped. The failures concern the two totality-debt counters, test-only module classification, and the DAE LOC review ledger. The classifier misses external child modules beneath an inline #[cfg(test)] module; added production assertions also need review. The DAE ledger reports core/total measurements of 17164/22571 lines. The separately completed workspace doctests pass 30 tests; 23 existing documentation examples are ignored. No gate ceiling is changed here.

The 28-row corpus check passes 27 rows and rejects the obsolete EL005 refusal pin for Rotational FirstGrounded, which now simulates and has all 46 channels high in the complete OMC comparison. Its pin now checks four states at 0, 0.01, and 0.02 seconds against independent numerical values. From the source gear ratio and inertias, the motor reflected onto inertia2 gives J = 0.1*10^2 + 2 = 12. With all four initial states zero, the reduced equations are 12*der(w2) = 100*sin(10*pi*t) - 10*w2 - 10000*(phi2-phi3) and 2*der(w3) = 10000*(phi2-phi3), with der(phi2)=w2 and der(phi3)=w3. A 60-decimal matrix exponential, augmenting the system with sine/cosine states, gives identical values with 80 and 100 terms. The largest Rumoca state error over the three samples is 8.91e-7. The new observations use the existing recorder’s initial and stepped tolerance formulas; the other 27 corpus rows are unchanged. The focused corpus check then passes with all 12 observations (verify-quick-firstgrounded-pin-2.log, 14.2 seconds). verify-quick-checkpoint-1.json and its logs in the campaign ledger retain the full checkpoint and follow-up validation. The full quick suite has not been repeated after the pin update; the rectifier and four gate failures remain open, and no baseline is promoted.

DAE module review

The 2026-09-11 production inventory uses SPEC_0043’s physical-line convention, including its test/generated exclusions. Relative to branch HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, core grows from 16,776 to 17,163 lines and wire from 5,292 to 5,407: 22,068 to 22,570 total, a 502-line increase. The mechanically derived ledger ceilings are 17,250 core, 5,500 wire, and 22,750 total. This acknowledges reviewed source size; no totality-debt ceiling or source-size review trigger is relaxed.

The derivative constructor proves complete functions, pure external bodies, the exact original-input prefix, role coverage, recursively filtered Real record fields, compact tangent shapes, priority uniqueness, and the issued predecessor chain before appending a link. First and higher derivatives use one insertion and signature-checking path. Consumers receive a borrowed view of those checked facts; no second annotation table is added downstream in DAE.

Wire stores target, input roles, predecessor identity, priority, and provenance. It does not serialize the derived order or tangent-prefix length: replay reconstructs them through the same constructors. Its pending traversal permits cross-function predecessor ordering and rejects a stalled or cyclic chain. This is distinct from semantic derivative selection in the structural phase. Review found no redundant validator, unchecked compatibility format, or second signature checker to remove.

The record traversal did redundantly reconstruct field indices and assert that lookups succeeded. A borrowed field iterator now carries each name and type together, eliminating both assertions while preserving order, nested Real filtering, and rejection of missing or extra tangent fields. The remaining production increment carries the construction and replay obligations above. The complete file inventory is retained as .git/multibody-campaign/dae-module-review-inventory.json.

Verification follow-up

All four workspace failures from the broad checkpoint are repaired. The totality scanner now follows Rust module items, including external descendants of inline test modules, while preserving any file also reachable through a production declaration. Files without a proved test-only path stay counted. Filesystem regressions reproduce both the missed test descendants and the unsafe exclusion of a dependency shared by production and test roots. The scope tests fail before the repair and pass afterward.

Production cleanup removes the actual added obligations separately: affine coefficients use a closed arithmetic enum, expression reads reuse the existing branded selector accessor, record traversal carries field names and types together, initialization keys distinguish Y and P directly, and matrix dependency construction produces the exact left/right coordinate maps as a pair. Structured initialization uses the existing domain iterator and carries its resolved scalar into body lowering. Totality ceilings are unchanged.

The affected libraries pass 582 tests; architecture passes 243 and repository gates pass 17. Workspace all-target/all-feature lint, formatting, and traversal checks pass. The subsequent complete workspace nextest run passes all 7294 tests with zero failures or skips, including the previously failing gates and the required MSL simulation regressions. These results are recorded in totality-construction-{libs,gates,lint,workspace}-1.log under the campaign ledger. The earlier workspace doctest result remains 30 passed and 23 ignored; doctests were not repeated for this private-API cleanup.

The fixed 20-model canary in target/msl/multibody-totality-canary, at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311 with worktree digest 65b0a8d54fae0879db4e3b2ee29b32da6e6073f6dfc006c1a18fb79602942703, has no phase, simulation-status, or comparison-band changes from multibody-prepared-causality-canary. All nine compared models remain high, including all 175 initialization channels; missing, skipped, excluded, nonidentifiable, and deviating counts are zero. The remaining eleven canary models retain their existing failures. The durable delta is totality-construction-canary-delta.json in the campaign ledger.

The full 566-model evidence remains the preceding broad checkpoint: 141 high, one actionable rectifier counterexample, and 17 existing exclusions. The full quick suite has not been repeated after these repairs, and full verification remains incomplete. Within that checkpoint, the 42 MultiBody examples retain 33 successful DAE constructions, 17 attempts reaching initialization, and 13 completed simulations with high trace parity. Counterexample investigation resumes before any additional breadth work.

The next diagnostic replays the rectifier’s exact checked SolveModel archive and freezes all solver coordinates and parameter values at t=0.002 and t=0.0594. Fresh reference-evaluator runtimes compare ordinary and coordinate-certified refresh at the same 1e-10 tolerance. Ordinary refresh retains neutral potentials of -1.35425e-6 V and 0.00297154 V; certified refresh reduces them to 1.491e-11 V and 1.155e-9 V. The corresponding maximum residuals fall from 1.91e-8 and 4.72e-9 to 4.26e-14. Both paths return success. This isolates a convergence-policy lead without advancing time or changing the stored inputs. A reduced public-observation regression and native runtime fix are still required; this diagnostic does not close the counterexample.

neutral-point-probe-2.log and neutral-point-probe-results.json retain the experiment. Its first setup attempt, neutral-point-probe-1.log, failed while decoding the DAE archive with InvalidArrayExtent, before any numerical evaluation. The successful follow-up uses the existing checked SolveModel replay API; the separate DAE archive-replay defect remains to be triaged.

Public observation convergence

The frozen rectifier evidence reduces to a two-variable source model:

model CoupledObservation
  parameter Real epsilon = 1e-6;
  output Real x(start=1);
  output Real y(start=1);
equation
  x*x + y*y = 2;
  x*x + (1 + epsilon)*y*y = 2 + epsilon + epsilon*1e-4*time;
end CoupledObservation;

Subtracting the equations gives y² = 1 + 1e-4*time and x² = 1 - 1e-4*time. Positive starts select the positive continuous branch. The regression proves that lowering retains a coupled two-variable projection block and checks both the public state observation and named getters at 0, 0.1, 0.5, and 1 second. Before the repair, all four combinations of BDF/RK and native/interpreted evaluation return x=1 at 0.1 seconds instead of 0.9999949999875. OpenModelica simulates the same source without warnings; its 12 CSV rows agree with the analytic solution within 1.619e-10. The checked source equations are correct, and the failure first appears when the public observation accepts a small equation residual despite a large coordinate error. This is a runtime convergence defect rather than a compiler equation defect or an integrator time-advance defect.

The linked FMI kernel now uses its existing coordinate-certified algebraic refresh for public observations, including each pass of the discrete/algebraic fixed point. It preserves the exact solver-vector width check and the frozen continuous-state prefix. This implements the observable-result preservation contract in SPEC_0038’s Internal Solver Boundary, under SPEC_0033 §§2–3 and §6a’s counterexample workflow. It changes neither the source model nor the comparison or numerical tolerances.

A second regression covers certification through causal back-substitution. For v = 1e6*i and i = 0.5e-6*v + 1e-11, the unique solution is i = 2e-11, v = 2e-5. Previously, the torn solver reported the zero vector settled because the current correction alone met tolerance, concealing the voltage error. The regression directly invokes the torn path and requires it to converge without a dense fallback. Certification now checks the undamped Newton candidate through the existing causal sweep and requires every recovered coordinate, as well as every tear coordinate, to meet its existing scaled correction tolerance.

Both fixes pass all 425 solver tests and all 456 core tests. Evidence is in observation-convergence-{red,green,solver}-1.log, observation-torn-certificate-red-1.log, and observation-convergence-core-2.log under the campaign ledger. The source-level regression is suite_core/public_observation_convergence.rs; the torn-path regression is runtime/projection/tests/certification.rs. The independent OMC receipt is observation-convergence-omc/analytic-comparison.json. The normal-budget originating-model comparison in target/msl/multibody-observation-rectifier, at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311 with worktree digest 8ffffe4abca43bb488dfccd9881270af6c4f84919cb8fe3f59e35e4e5a7f6aae, completes simulation in 1.276 seconds. It compares all 429 channels and all 429 initialization values, with zero missing, skipped, excluded, or nonidentifiable traces. Seventeen formerly severe channels now agree, including the neutral voltage; its maximum reference difference is approximately 6.337e-9 V. However, three linked rootMeanSquareVoltage channels remain deviating: mean.y, mean.y_last, and sqrt1.u. Their mean absolute error is 2.018, with a maximum of 10.358. Thus 426 channels agree, but the model remains non-high and the counterexample remains open. The focused command’s zero exit status does not override its comparison result. The complete cohort still requires new validation; unrelated capability work stays paused.

The fixed 20-model canary in target/msl/multibody-observation-canary, at the same HEAD with worktree digest 6a91166f0e27748f4711fce1a6bfe8986f5c41ec6d586dfe4cd7e61012b48bdd, preserves every phase, simulation status, and comparison band from multibody-totality-canary. All nine compared models remain high, including all 175 initialization channels; missing, skipped, excluded, nonidentifiable, and deviating counts are zero. Its eleven existing failures remain visible. The durable delta is observation-convergence-canary-delta.json.

The remaining RMS error is present in the pre-repair trace as well. Its three channels alias the mean block’s held output: der(x)=u, with squared voltage as the input, and y_last=f*pre(x) followed by reinit(x,0) at each 300 Hz sample. The first diagnostic preserves the before/after Rumoca and OMC sample values in observation-rms-initial-triage.json. Integral endpoint accuracy, event pre-state sampling, and reset handling remain competing explanations; none is yet established as the next defect.

BDF accepted state and sampled integrals

The RMS investigation finds a discontinuity between the recorded integral one representable time before a sample and the pre(x) value used at that sample. The time gap is too small for the bounded derivative to explain the state gap. A reduced model keeps only der(x) = 50000 + 20000*sin(2*pi*300*time) and a 300 Hz sample that stores pre(x) and resets x to zero. On the first tick, BDF reports left x=166.6647700240386 and sampled pre(x)=166.66516325266724, across a time gap of 4.337e-19 seconds. The same source passes on the RK host. OpenModelica runs all 30 resets with exactly matching event-entry and sampled values; its maximum analytic integral error is 2.182e-4, at the same 1e-6 simulation tolerance.

A direct Diffsol regression removes Modelica, FMI, events, and Rumoca’s sampler entirely. For x'=x, its first accepted state is 1.0001414213562374 at t=0.00014142135623730943, while native interpolation at that exact time is 1.000141438235889. This proves that the first divergent owner is the numerical dependency’s accepted-state publication. Diffsol updates its backward differences with the Newton correction but publishes the uncorrected predictor as state.y; an event reset then restarts from that predictor. Rumoca’s endpoint-copy special case hid the disagreement from the common host’s endpoint consistency check (SPEC_0038 / SPEC_0044 §6).

The dependency patch publishes column zero of the updated backward differences as the accepted state. Rumoca removes its endpoint overriding copies and uses the native extension throughout the accepted interval. This follows SPEC_0033 §§2–3’s first-owner rule and preserves the MLS §8.3.6 sampled-state and reset semantics. The direct regression and both source-level host variants pass, followed by all 458 core tests. Evidence is in native-bdf-endpoint-{red,green}-1.log, sampled-integral-{red,green}-1.log, sampled-integral-omc/analytic-comparison.json, and bdf-endpoint-core-1.log.

vendor/diffsol/RUMOCA_PATCH.md records the immutable upstream revision, published archive checksum, one-line source change, license, and release limitation. The archive audit verifies that every original package file is present, with only src/ode_solver/bdf.rs changed. A workspace patch cannot repair registry-published downstream crates: registry publication still needs an upstream release carrying this fix or a published maintained dependency. No upstream submission has been made. Origin, canary, and complete-cohort comparison remain pending; the rectifier counterexample is still open.

The subsequent originating-model run, target/msl/multibody-bdf-endpoint-rectifier, completes in 1.645 seconds and compares all 429 channels, with all 429 initialization values high and no missing, skipped, excluded, or nonidentifiable trace. At HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, worktree digest b8008171426889515029f6c7d99eb9c4b13a585732c126d02744d2797e0941c4, 426 channels remain high and the same three RMS channels remain deviating (mean absolute error 2.01643, maximum 13.68547). The integral’s recorded left value and sampled pre(x) now agree within 2.558e-13, so the endpoint defect is repaired, but it does not close the RMS counterexample.

The fixed 20-model canary in target/msl/multibody-bdf-endpoint-canary binds the same worktree digest and preserves every phase, simulation status, and comparison band from multibody-observation-canary. All nine compared models remain high, including all 175 initialization channels; missing, skipped, excluded, nonidentifiable, and deviating counts are zero. Eleven existing failures remain visible. The delta is bdf-endpoint-canary-delta.json.

The next direct kernel regression finds a separate derivative-coordinate defect. At a known scheduled boundary, continuous_eval_time moves physical time backward by twice the state tolerance. For the smooth derivative 1e6*time at t=1, the fixture returns 999800 instead of 1000000. It agrees before the boundary. The repair evaluates the scheduled left limit at the adjacent representable coordinate, as the scheduled event-entry snapshot already does, and leaves numerical root-bracket probing unchanged. The regression fails before the repair and all 426 solver tests pass afterward (scheduled-derivative-coordinate-red-1.log and scheduled-derivative-solver-1.log). Broader and originating-model validation of this additional repair remain pending.

The additional repair passes all 458 core tests and all-target/all-feature Clippy for the solver and core packages. Its originating-model run in target/msl/multibody-scheduled-derivative-rectifier completes in 1.517 seconds, with 426 of 429 channels high, all 429 initialization channels high, and no missing, skipped, excluded, or nonidentifiable trace. The three RMS channels still deviate; their mean absolute error increases to 2.55457 and their maximum error is 14.58486. This is a repaired derivative-coordinate defect, not closure of the RMS counterexample.

The fixed 20-model canary in target/msl/multibody-scheduled-derivative-canary retains every phase, simulation status, and comparison band from multibody-bdf-endpoint-canary. At HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, its worktree digest is fd999cb8aa11cfb4b80bcc096e5fa9fd7ec336cfdf3e3f370873b20c7c1791f3. All nine compared models and all 175 initialization channels remain high; missing, skipped, excluded, nonidentifiable, and deviating counts are zero. The durable delta is scheduled-derivative-canary-delta.json.

The earlier rectifier trace also shows firing transitions hundreds of nanoseconds after the OMC transitions. The source controller composes voltage zero crossings, Modelica.Blocks.Logical.Timer, a gain, and a strict threshold comparison. Root localization and recognition of the resulting timer deadlines remain candidates for the residual averaging error. They require a reduced proof before another production change. A new complete quick run is next to measure the combined runtime changes across the full cohort and workspace.

Verification checkpoint after observation and BDF repairs

The complete multibody-verify-quick-checkpoint-2 archive binds HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311 and working-tree digest ae29d3016cdfffa208e9114e7c0ba58b4e294672c0ac756280ab53dec6a754a0. The 566-model comparison retains all 141 previously strict-high models, with 142 compared, 16 existing policy exclusions, zero missing traces, and zero nonidentifiable traces. The rectifier remains the sole actionable model: 426 channels are high and the same three RMS channels deviate. All 14,550 compared initialization channels remain high. No baseline was promoted.

The 42 MultiBody examples retain 33 balanced DAEs, 17 initialization attempts, 16 successful initializations, and 13 strict-high completed simulations. All 5754 compared channels are high, with no missing or excluded comparisons. The 29 failures retain their previous phase classifications. Outside MultiBody, the already excluded CenterTap2Pulse RLV_Characteristic newly times out; the complete phase delta also records three changed failure diagnostics. This execution loss is not hidden by the unchanged strict-high band table.

verify quick completed in 2201.2 seconds, failing only its MSL soundness stage. Workspace lint, all 28 corpus rows, 243 architecture tests, 17 repository gates, and 7303 workspace tests pass. Documentation tests have 30 passes, zero failures, and 23 ignored tests. These results precede the next event relation repair; they do not close the remaining counterexample.

Rectifier event relation, traced against OMC

The next investigation follows the original rectifier through its first switching event, using the archived full-cohort traces and the source of Signal2mPulse, Logical.Timer, and Logical.Greater:

1. At 50 Hz and a 30-degree firing angle, the threshold is 1/6 and the timer gain is 100. The first negative timer must stop when its voltage changes sign near 1/600 seconds. OMC’s settled row sets the timer and its firing output to zero. Rumoca’s row at 0.0016673391524120351 also sets the timer and greaterNegative[1].u1 to zero, but leaves greaterNegative[1].y=1 against u2=1/6. The false pulse persists until 0.0016676665409000395.
2. The retained Solve artifact has the correct root expression, u2-u1, at root 19, with an AlgebraicDependent relation target P359. The visible Boolean binding and its output program both read that same P359. This rules out an inverted comparison or a mismatched output slot.
3. The completed-step callback selects the crossed root’s new side. Event iteration then turns off the upstream timer and projects u1=0, reversing root 19’s sign. The old runtime unconditionally excludes the selected root from relation refresh and reapplies its override on every pass. It can therefore report convergence with a false relation equation. This violates the mixed-equation event solve in MLS Appendix B.
4. A reduced source containing only a sine crossing, a conditional timer, and a threshold reproduces the pulse with both BDF and RK at unchanged 1e-6 tolerances. OMC keeps fire=0 on all 104 rows of the identical source. A runtime-only two-root test reproduces the stale Boolean without compiler lowering or numerical integration.

Event iteration now releases a selected side when the projected root becomes strictly nonzero on the opposite side. The surviving selections pass through post-commit canonicalization, preventing a released selection from returning. The exact-zero control retains its selected side. Post-commit refresh still preserves the separate frozen-parameter relation contract. The earlier test that asserted unconditional override authority now checks both exact-zero retention and release after a strict sign reversal.

The runtime regression and all 427 solver tests pass. Both reduced source variants pass, as do all 460 core integration tests. These focused checks do not establish closure of the original RMS deviation; its next OMC comparison and the fixed canary delta are still required. The private receipts retain the first-event comparison, red/green logs, and source-identical OMC CSV hashes.

All-target/all-feature Clippy for the solver and core crate also passes. The post-commit fixture was extracted to satisfy the function-size limit; all 11 event-iteration tests pass afterward, and the formatting check passes.

The fixed 20-model canary is now complete at target/msl/multibody-event-relation-canary: all phase and simulation statuses match the preceding canary. The same nine compared traces are strict-high, with all 175 initial channels high and no missing, skipped, excluded, or nonidentifiable candidates. This is a focused regression result, not a cohort coverage number.

The originating-model gate at target/msl/multibody-event-relation-rectifier failed before compilation: the model worker exceeded the unchanged 60-second source-root loading budget. It attempted no simulations and reports parity unmeasured. That failed attempt is retained. A separate bounded one-shot source diagnostic then compiled and simulated the original model. Its first negative firing pulse is gone, but the comparator still finds the same three RMS deviations among 429 channels; all 429 initial channels remain high. The source diagnostic does not replace the failed gate. An attempted replay of the older DAE JSON was also rejected at deserialization with array extent must be a nonnegative literal Integer; that artifact round-trip limitation remains untriaged.

Rectifier switching-time counterexample

The repaired relation did not materially change the RMS values, so it was a separate observable defect rather than the explanation for their remaining error. The next source-to-OMC trace follows the first positive firing pulse:

The 0.890986-microsecond delay keeps the old circuit branch active across a voltage-squared jump of approximately 54449.996. Its missing integral, multiplied by the mean block’s frequency of 300, predicts an error of -14.55426; the observed error is -14.58486. This calculation links root timing to the accumulated error without adjusting any comparison threshold.

A reduced source integrates if sin(100*(time-1/300)) > 0 then 10000 else 0 over 0 to 0.01 seconds. Its exact solution is x(t)=10000*max(0,t-1/300). Both Rumoca integrators violate the existing 1e-6 state-error scale around the event; BDF still reports zero at 0.003333500000000001, where the exact integral is 0.001666666666674116. OMC’s source-identical 54-row trace has maximum absolute error about 1e-8. The private root-integral-omc receipt retains the source, regression fixture, CSV, and hashes.

The current host bisects to the session’s time-width bound and then applies the upper bracket endpoint. Its default constructor derives that time bound from the absolute state tolerance. The bisector satisfies its present bracket contract; these results expose a numerical accuracy problem across the derivative jump, rather than proving an inverted root or bad interpolation inside either integrator. FMI assigns this localization to the common host before the completed-step callback (FMI 3.0.2).

A bracketed secant prototype at unchanged tolerances repaired BDF but still failed the RK reduced case. It was removed from production, with its patch and failure log retained privately. The reduced counterexample remains open; no root-policy change, tolerance adjustment, baseline promotion, or new full-model support claim follows from this experiment.

Root-time policy construction

The next proof isolates default option construction. Two metamorphic tests fail without changing a model: changing the absolute state-error scale from 1e-9 to 1e-3 changes the default root-location duration by six orders of magnitude, while multiplying the time coordinates by 1024 leaves that duration unchanged. The constructor had supplied the absolute state tolerance where a time-resolution policy was required.

Default batch and live construction now obtain that duration from the existing accepted_step_roundoff(start_time, scan_resolution) policy, bounded by the scan resolution. The helper that accepted a state tolerance has been removed. The proposed contract is recorded in SPEC_0038 and SPEC_0044 §6: state-unit changes leave time resolution unchanged, and time-unit changes rescale it. This changes internal default root accuracy through dimensional ownership; requested state-error settings, comparator thresholds, explicit root-location options, scan cadence, and the host time-roundoff formula retain their previous definitions. The bisection implementation is unchanged.

Both metamorphic tests and all 429 solver tests pass. The original analytic switched-integral regression now passes for both BDF and RK with its unchanged 1e-6 settings. Its OMC source and trace remain bound by the earlier receipt.

The first broader core run passed 461 tests and exposed one timestamp snapshot in the function-assertion message test. The diagnostic payload remains f rejects: u=2 y=6. Its reported time moves from 0.095894271 to 0.095894015, closer to the analytic crossing ln(4/3)/3 = 0.09589402415059362. That test now checks the typed component assertion, its exact message, and the analytic state at the reported time against the requested state-error scale. The focused test passes. The complete core suite passes 462 tests, and the solver and compiler all-target, all-feature Clippy checks pass.

The original-model run (multibody-root-time-policy-rectifier) compares all 429 channels: 426 high, three deviating, zero severe, and all initial channels high. The same RMS channels remain counterexamples. Their mean absolute error falls from 2.55 to 0.66, with maximum error 5.44. The fixed 20-model canary (multibody-root-time-policy-canary) retains every status: nine compared models remain entirely high with 175 high initial channels, and eleven pre-existing failures remain visible. No baseline is promoted.

The remaining event rows narrow the next investigation. The positive timer’s entry is accurate at 0.0016666666666668587 s. At 0.0033333333333335708 s its comparator input is 0.1666666666666712, above 1/6, yet the Boolean remains false until 0.0033333651543144957 s. A separate phase changes voltage sign near 0.005 s but settles its Boolean only at 0.005000332031542508 s. These are observations to trace through root evaluation and event iteration, not grounds for a comparator exception or a claim of closure.

A consumed sample tick backdates a state event

The first timer discrepancy is now reduced to SampledTimer: a sine crosses zero at 1/600 s, a when stores its entry time, and a timer enables an integrator after another 1/600 s. An independent sample(1/300, 1/300) counter provides the coincident scheduled owner. MLS Appendix B and SPEC_0022 SIM-001, SIM-008, and SIM-010 require the relation to settle at its event coordinate without executing the sample owner again.

Before the fix, BDF reaches the timer root at 0.003333333333333348 s with an accurate stored entry time, but the kernel applies the event at the consumed tick, 0.0033333333333333335 s. The relation becomes false there. Its next accepted step reaches 0.003334333333333348 s before enabling the integrator, losing 0.01 from the exact integral. This excludes integration error and timer lowering as the first cause: the runtime probe directly identifies the coordinate change in event_update_application_time.

OMC completes the identical source with 61 rows and exactly three sample activations. Its maximum absolute error against the analytic integral is 1.005e-6, compared with Rumoca’s 0.01 loss at the delayed event. OMC’s located event row is just beyond the reduced test’s 1e-6 analytic bound; the oracle is evidence for the trajectory, not an exact-arithmetic reference.

The application-time selector now consumes the existing three-way StateTimeCoincidence value. Only an unconsumed clock selects its semantic tick; a consumed clock retains the component coordinate while the existing row filter prevents replay. No root-search policy, state-error setting, or comparison threshold changes. The regression checks the analytic integral for every trace row under BDF and RK and checks exactly three sample activations. Both reduced solver cases pass, as do all 429 solver tests, all 464 core tests, formatting, and all-target, all-feature Clippy for the solver and compiler.

The original-model run multibody-sampled-timer-rectifier now compares all 429 channels high, with all 429 initial channels high and zero missing, skipped, excluded, or nonidentifiable comparisons. This closes the three RMS counterexamples for this model. RMS mean absolute error is 0.0121 and maximum absolute error is 0.0953. The first positive firing settles at 0.0033333333333335708 s; the later phase-three voltage transition settles at 0.005000000000000596 s. No comparison policy changed.

The fixed 20-model canary multibody-sampled-timer-canary retains every phase and simulation status and every channel band: nine compared models remain entirely high, with 175 high initial channels, zero missing/skipped/excluded/ nonidentifiable comparisons, and eleven pre-existing failures. Both runs bind HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311 and dirty-tree digest c9243d92d1ce6253763bfd4f3f93b8675ba2cd81be293cffa371573bfaf39ee8. These focused results are not a cohort coverage claim. The original model remains in the next complete 566-model comparison; no baseline is promoted.

Complete cohort after the consumed-tick fix

multibody-sampled-timer-full completes the full 566-model gate with exit 0 at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, dirty-tree digest 5364e55e1d435913b77da2a673065ab92736dcd208f90a140e42e19ea04eecb2. All 142 compared models are strict-high, with zero deviating or severe channels, zero missing traces, 15 skipped models under the existing policy exclusions, and zero nonidentifiable traces. All 14,550 initial channels are high. The 141 previously strict-high models remain strict-high; the repaired rectifier supplies the one additional model. The 15 existing minor channels across three other models retain their previous classifications.

MultiBody’s 42-example scope is unchanged: 33 balanced DAE constructions, 17 initialization attempts, 16 successful initializations, and 13 compared strict-high simulations containing 5,754 channels. The other 29 examples remain failures; this runtime fix adds no MultiBody breadth claim.

Raw execution falls from 158 completions to 157. The previously completing, policy-excluded RectifierBridge2mPulse.HalfControlledBridge2mPulse now reports EX001: BDF step-size exhaustion at 0.013333333333334703 s. This is an execution regression retained for separate triage, despite having no effect on the strict-high count. Four other already-failing models change failure classification; their complete phase delta is retained in the campaign receipt. There was no retry, policy change, or baseline promotion.

This milestone reruns the focused solver/core suites, Clippy, the fixed canary, and full MSL comparison. It does not claim a new verify quick or verify full pass; those complete workflows remain required before finalization.

BDF restart steps below its configured minimum

The excluded half-controlled rectifier’s execution regression is independently reproduced from its source: it fails at 0.013333333333334703 s. The retained OMC trace has smooth load voltage through this event, as does the prior completing Rumoca trace. A temporary derivative probe confirms smooth runtime values: approximately 269.443602 V and 72599.854757 V² for the first two integrals. This excludes a new voltage discontinuity as the first cause.

At the located root, the RMS integral is only 3.31e-11. Diffsol’s initial-step heuristic chooses approximately 4.56e-16 s, below its configured 1e-13 s minimum. BDF accepts two steps and then raises StepSizeTooSmall while increasing the step following successful solves. A direct Diffsol test with the constant derivative 72599.85 and initial value 3.31e-11 reproduces the exact failure time without Modelica, FMI, clocks, roots, or algebraic refresh. A second constant-derivative test fails when continuing after a hard stop whose interval is shorter than the configured minimum.

The numerical owner now applies its configured minimum at the beginning of each new BDF attempt, then reapplies any closer hard stop. The difference table is rescaled to the actual new step size. Error-driven reductions retain the existing minimum-step failure; no model, state-error setting, comparator policy, or FMI adapter workaround changes. The local dependency patch remains confined to vendor/diffsol/src/ode_solver/bdf.rs, with provenance recorded in vendor/diffsol/RUMOCA_PATCH.md. Current upstream commit a33f02a4952c6837979754cab92eef70763a2f41 retains the original interaction; no upstream submission has been made.

Both direct regressions pass forward and backward. A separate stiff problem still fails with the expected minimum-step error when its requested accuracy cannot be resolved, and the prior accepted-endpoint regression remains green. All 464 core tests, formatting, and all-target/all-feature Clippy for the Diffsol adapter and compiler pass. The temporary probes have been removed.

The fixed 20-model canary multibody-bdf-minimum-step-canary retains every phase and simulation status and every model band: nine compared models high, 175 high initial channels, zero missing/skipped/excluded/nonidentifiable comparisons, and eleven existing failures.

The complete multibody-bdf-minimum-step-full gate passes with 566 targets, 142 compared strict-high models, zero deviating or severe channels, all 14,550 initial channels high, zero missing/nonidentifiable traces, and 16 skipped models under existing policy exclusions. All 142 previously high models remain high. The half-controlled bridge completes in 1.75 s, restoring the raw execution count from 157 to 158; its existing comparator exclusion remains and it receives no parity credit. Two already-failing electrical models now report timeout instead of their previous numerical failures.

Both runs bind HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311 and dirty-tree digest 807587dd97ad47a9c3466919fe4be5ae45ba9dc251466ef62f7e693055900a49. MultiBody remains at 33 balanced DAE constructions, 16 successful initializations, and 13 strict-high simulations out of its 42 examples, with 5,754 compared channels. No baseline is promoted. Complete verify quick and verify full workflows remain required before finalization.

PointGravity: initialization guesses evaluated too early

With the execution regression closed and zero actionable trace deviations, the next original model is MultiBody.Examples.Elementary.PointGravity. Its fixed body positions are {0,0.6,0} and {0.6,0.6,0}. The Body equations alias those positions to the integrated frame_a.r_0 coordinates and compute point gravity from the radius. Fresh OMC simulation reaches 5 s with 502 rows; body two starts at {0.6,0.6,0} with gravity {-0.9820927516479827,-0.9820927516479827,0}.

The retained Rumoca source/Flat/DAE/Solve diagnostic fails during initialization with non-finite gravity. The declared body positions have the correct starts, but the integrated position guesses are zero. Orientation matrices are already valid. The initialization constraints retain the fixed values; the failure occurs before those constraints can settle the state coordinates.

The reduced source has a = b, b(start=0.6,fixed=true), and der(a)=-1/a. That control passes both BDF and RK. Extracting the equivalent algebraic g=-1/a; der(a)=g makes both fail with a non-finite g before initialization. OMC initializes the exact reduced source with a=b=0.6 and g=-1.6666666667 and completes its trajectory. This distinguishes premature algebraic evaluation from lost fixed attributes or a derivative-solver failure.

SolveMeKernel::initialization_solver_y called the ordinary continuous refresh while collecting its guesses, before settle_initialization_system. It now copies the retained guess and current state prefix without evaluating ordinary algebraics. The existing joint initialization solve remains the owner of transferred pins, dependent starts, parameters, and manifolds under MLS §8.6 and SPEC_0043’s initialization construction contract. No IR contract, pin classification, tolerance, or source-specific branch changes.

The reduced BDF/RK cases, all 429 solver tests, and all 466 core tests pass. Formatting, whitespace checks, and all-target/all-feature Clippy for the solver and compiler pass. The original-model gate multibody-initial-alias-point-gravity reaches 5 s with all 129 compared channels high, including all initial channels, and no missing, skipped, excluded, or nonidentifiable traces. Body two’s initial position and gravity match the fresh OMC reference values above.

The fixed 20-model multibody-initial-alias-canary retains every phase, simulation status, and comparison band from multibody-bdf-minimum-step-canary. Its nine compared models and 175 initial channels remain high, with no missing, skipped, excluded, nonidentifiable, or deviating comparisons; eleven existing failures remain visible. Both focused runs bind HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311 and dirty-tree digest 5252b059e681b90c9158b7e30359caf205e6b7edfa322983a702c31a1e172919.

The first complete-cohort attempt aborted with ENOSPC while writing worker results. It produced no aggregate or comparator result and has no parity claim. Removing approximately 50 GB of regenerable parser and incremental compiler caches preserved all source, traces, and verification evidence. A fresh full-cohort run used the same source, targets, budgets, and comparator.

That complete gate, multibody-initial-alias-full-after-disk-recovery, passes with 566 targets and 143 compared strict-high models. All 142 previously high models remain high; PointGravity is the only new high model. There are zero deviating or severe channels, zero missing or nonidentifiable traces, and 16 skipped models under the unchanged policy exclusions. All 14,679 initial channels are high. This run binds HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311 and dirty-tree digest 1235402343f0ecdae5a1395886d0e6a72f4f95da0dad21f7841745b60dd5367a.

MultiBody now has 33 balanced DAE constructions, 17 successful initializations, and 14 compared strict-high simulations out of its 42 examples. Its 5,883 compared channels have zero deviations, with no missing, skipped, excluded, or nonidentifiable comparisons. The other 28 examples retain their failures: two Flatten timeouts, six DAE construction failures, ten Solve timeouts, six structural failures, three simulation timeouts, and one Instantiate failure.

One already-failing Media model, SolveOneNonlinearEquation.Inverse_sh_TX, changes from DAE parameter-binding error ED019 to Flatten record-layout error EF015 after cache regeneration. Its simulation band remains absent; this phase variation is recorded but its cause is not yet established. No previous successful simulation is lost. No baseline is promoted. Complete verify quick and verify full workflows remain required before finalization.

Complete verification checkpoint after PointGravity

cargo xtask verify quick passes under the fixed four-job concurrency budget. The checkpoint includes workspace lint, full MSL parity, all 28 pinned corpus rows, 243 architecture tests, 17 repository tests, all 7,317 workspace tests with none skipped, and 30 passing documentation tests with 23 existing ignored examples. The complete log and timing report are retained in the campaign.

The MSL snapshot is multibody-verify-quick-checkpoint-3, at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311 and dirty-tree digest 18569abeb3ac1c904299276276a151693a099b323cd8731cc71a75f6076e24d0. Every one of its 566 phase/simulation outcomes and comparison bands is unchanged from multibody-initial-alias-full-after-disk-recovery. It retains 143 compared strict-high models, zero deviating channels, zero missing or nonidentifiable traces, and 16 skipped models under the existing exclusions. MultiBody retains 14 compared strict-high examples and all 5,883 channels high, with no skipped or missing comparisons. This is a successful complete verify quick checkpoint; verify full and the remaining MultiBody scope are still outstanding. No baseline is promoted and no PR is opened.

Surfaces: forwarded replaceable function interface (in progress)

The original MultiBody.Examples.Elementary.Surfaces fails in Instantiate with EI027. PipeWithScalarField declares a replaceable colorMap alias for Colors.ColorMaps.jet, constrained by Interfaces.partialColorMap, and forwards that alias to Advanced.PipeWithScalarField. jet explicitly extends the interface. The retained OMC baseline completes the original 5-second model with 502 rows and 336 channels.

The reduced source reuses the existing function-redeclaration fixture: PartialF has input x and output y; Double and Triple implement it; Consumer defaults to Double; and Wrapper forwards its F = Triple alias into Consumer. A non-replaceable forwarding alias compiles and returns 6. Adding replaceable and constrainedby PartialF reproduces EI027. Both exact sources, extracted from the Rust fixture strings, compile and simulate in OMC with y = c.y = 6 at every row through 0.2 s. This is a confirmed reduced failure, not a repaired capability.

A temporary predicate trace identifies the same first rejection in the reduced and original models. For the reduced case, replacement declaration DefId(110) has explicit constraining DefId(99), exactly the required PartialF DefId(99). For the original, colorMap DefId(571) has constraining DefId(391), exactly the required partialColorMap DefId(391). Both pass the earlier subtype/class checks and fail class_flags_compatible when the referenced declaration’s replaceable flag is compared with the standalone constraint declaration. The caller is type_overrides::component_redeclare_validation, reached through resolve_component_nested_type_overrides. The tracing has been removed.

The repair must distinguish a referenced class’s effective constraining interface from its declaration flags, under MLS §§6.3, 6.6, and 7.3. It must also preserve validation of the default implementation, constraining modifiers, and incompatible signatures. Merely replacing the source identity with the named constraint could bypass a necessary default-type check. These negative obligations still need reduced evidence before selecting the production change. No compiler repair or additional model pass is claimed; the newly added replaceable forwarding regression is intentionally failing after the successful verification checkpoint above.

The candidate repair now constructs callable reference interfaces in inheritance/function_interfaces.rs, keyed by resolved declaration IDs. A replaceable reference exposes its constraint only after its actual/default function satisfies that constraint. Short aliases follow the referenced interface and apply their binding modifiers through the existing inheritance helpers. Effective inherited public components feed the existing member and ordered function-signature comparator. The ordinary declaration flag checker is unchanged; component and extends function redeclarations share the new reference constructor.

The expanded pre-fix regression suite records three valid failures: explicit constraints, implicit constraints, and a bound additional input. The candidate passes all twelve tests and the complete 473-test core suite at that checkpoint. OMC accepts the exact implicit and bound-input variants, with four rows through 0.2 s and y = c.y = 6 throughout. A further extends-forwarding regression exposes the second caller of the old declaration comparison. After routing that caller through the reference constructor, all fifteen focused tests pass, including alias chains and rejection of an impure default under a pure constraint. The impure fixture explicitly declares its alias impure so that it exercises compatibility rather than Resolve’s earlier purity rejection.

These are candidate-level regression results. Original Surfaces, canary, complete-cohort, and final verification evidence for this repair are still pending; no additional MSL parity count or completed capability is claimed.

The final core run passes all 476 tests; the combined contract and instantiation run passes 689 and 215 tests respectively, with two existing ignored documentation examples. All-target/all-feature instantiate Clippy, formatting, and whitespace checks pass. The normal-budget originating run multibody-function-interface-surfaces advances through Instantiate, Typecheck, and Flatten in 1.67 seconds total, then stops in ToDae with ED019: pipeWithScalarField.pipe.colorMapData evaluation expects Integer but receives Array. No simulation or comparator runs, so this originating attempt has unmeasured parity and exits 1. The original EI027 rejection is cleared; the model is not yet supported.

The fixed multibody-function-interface-canary passes with all twenty phase, simulation, and band outcomes unchanged from multibody-initial-alias-canary. All nine compared models and 175 initial channels remain high, with zero missing, skipped, excluded, nonidentifiable, or deviating comparisons. Its receipt binds HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311 and dirty-tree digest 8ed2c9f142ab75d617e70116f2b487266349061ba8da3f53665615ec342dccc4. The subsequent complete multibody-function-interface-full cohort passes at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, with dirty-tree digest 7a606f2d678da165d65656988e7c7c6d2c5f1a2a79566485361b2439d5cecf46. All 566 targets and all model comparison bands are unchanged from multibody-verify-quick-checkpoint-3: 143 compared high-agreement models, 16 existing policy exclusions, zero missing or nonidentifiable traces, and zero deviating or severe channels. Across the full cohort, 18 channels retain the comparator’s minor band; the model bands remain high. The only phase change is Surfaces advancing from Instantiate EI027 to ToDae ED019. MultiBody retains 14 compared high models and 5,883 high trajectory and initialization channels, with no minor, deviating, missing, skipped, excluded, or nonidentifiable comparisons. Its stage counts are now 42 instantiated, 40 flattened, 33 balanced DAEs, and 17 successful solver constructions and initializations. This completes the function-interface cohort check, not verify full or support for Surfaces.

Surfaces: constant matrix-prefix indexing

The next original failure is the constant binding of pipeWithScalarField.pipe.colorMapData. The retained Flat artifact in .git/multibody-campaign/surfaces-slice-diagnostic/ir-flat.json has SHA256 275ae123228370df4599105a8b7b647004c2d5e593c45a0858d443aedf86fe19. Binding instance 626 has shape [32,3]; its call references function instance 2 and passes scalar instance 625, whose binding is Integer 32. The selected body retains jet’s final assignment colorMap := cm[1:n_colors, :], with the source range and colon intact. This rejects the competing explanations of a wrong forwarded body or an array-valued n_colors argument.

The first responsible evaluator is rumoca-eval-flat::constant::function_eval::apply_single_subscript: it previously required every expression subscript to yield one Integer. The range correctly evaluates to an array of indices, which that implementation rejects. MLS §10.5 and SPEC_0022 ARR-005/024/025 govern this indexing; SPEC_0029 §5 assigns Flat constant evaluation to the evaluator crate.

The reduced function_parameter_slicing fixture initializes a 3-by-2 matrix with distinct coordinates and returns cm[1:n, :]. Before the repair, that case reproduces ED019; an explicit matrix and a scalar column-selection control already pass. OMC accepts all three exact sources and emits the expected constant parameter values. It eliminates their constant observables from CSV, so the oracle comparison uses generated initialization XML and does not claim compared trajectory channels. Source and XML receipts are in .git/multibody-campaign/surfaces-slicing-omc.

The candidate shared subscript evaluator replaces the ordinary-expression and function-body scalar-only read implementations. It evaluates selectors in the caller’s scope, preserves the order and multiplicity of vector indices, and bounds-checks each selected coordinate. All three end-to-end regressions now pass, as do all 140 evaluator tests, including retained versus removed axes, repeated indices, empty selections, and invalid signed bounds. Broader core, lint, original-model, and canary validation was pending at that point.

The full core suite subsequently passes all 479 tests; all-target/all-feature evaluator Clippy and compiler-core Clippy pass. The original normal-budget run multibody-constant-slice-surfaces clears the Integer/Array mismatch and now fails the retained bounds check: colorMapData attempts index 6 in an array of size 5. It still stops in ToDae ED019, with no simulation or comparison. jet constructs its matrix from five block rows, whose scalar row counts sum to 32 for this input. The next investigation must distinguish block matrix concatenation from literal array nesting; bypassing the new bounds check would hide the incorrect intermediate shape.

The fixed multibody-constant-slice-canary passes at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311 and dirty-tree digest 6bed78a66188b7f3de4eec331a66b386708f56dd27fa1470f8d905532a1d1e74. All twenty phase, simulation, and band outcomes are unchanged from multibody-function-interface-canary. Its nine compared models, 175 trajectory channels, and 175 initial channels are all high, with zero missing, skipped, excluded, nonidentifiable, minor, or deviating comparisons. The delta receipt is .git/multibody-campaign/surfaces-slice-canary-delta.json. This completes Tier 1 for slice reads; no further MultiBody model pass is claimed.

Surfaces: preserve and evaluate bracket concatenation

The next reduced fixture, function_matrix_construction, builds a 3-by-2 matrix from two block rows: two length-2 vectors followed by two length-1 vectors. Returning cm[1:3,:] fails with index 3 in size 2 before the repair; an explicit scalar matrix control passes. The original retained jet Flat body likewise contains five bracket block rows, and both Flat constant evaluators previously ignored the constructor flag and merely nested each evaluated operand. This accounts for the original five-row intermediate.

A second regression exposes an earlier representational defect. After removing only provenance from its comparison, the parser produces identical constructor structure for [[11,12],[21,22]] and [11,12;21,22]. OMC accepts both and records dimensions 1-by-4 and 2-by-2 respectively. It also accepts the exact block fixture and its explicit control, with all six expected parameter coordinates. The pre-fix suite has one pass and two failures in surfaces-matrix-construction-red-1.log; the four source and XML receipts are under .git/multibody-campaign/surfaces-matrix-omc. These are parameter-value comparisons, not complete MSL trace results.

The candidate replaces the ambiguous AST/Flat Boolean with the shared ArrayConstructor enum: element construction, horizontal concatenation, or vertical concatenation. Parse records the source separator; rewrites preserve it; expanded comprehensions and materialized arrays remain element constructors. Fully known operand extents use the common checked-dimension operation named in SPEC_0041. DAE construction selects its existing checked promoted-concatenation builtin directly, removing nested-row guessing and its call-discovery workaround. A shared Flat value constructor promotes operands and concatenates along the recorded axis. AST display preserves that same operation. No old Boolean representation or fallback reader remains.

All five focused regressions pass in surfaces-matrix-constructor-candidate-3.log, including nested bracket blocks and horizontal concatenation of explicit column matrices. This is candidate evidence only: affected library suites are running, and complete core, lint, original Surfaces, and canary validation remain pending. The full-cohort MultiBody count remains the preceding fourteen-model result.

The affected library/integration run then passes 1,681 tests and one doctest (with nine existing ignored doctests), and the complete compiler-core suite passes all 484 tests. A private borrowed dimension scope repairs the nested comprehension shape regression exposed by the library run: lexical scalar indices have known rank without an invented constant value, and an outer same-named parameter cannot supply a varying local dimension.

Further review adds two reduced failures in surfaces-matrix-scope-red-2.log: a resolved component with an explicit InstanceId is incorrectly shadowed by a same-spelled comprehension index, and an expression subscript unnecessarily requires a compile-time integer value to determine its scalar result shape. The repair recognizes lexical indices only when the reference has no concrete instance identity, using the same predicate for shape inference and value-dependency inspection. Subscript projection uses the index expression’s rank: scalar removes the axis, vector retains its length, and higher rank remains unsupported. Actual constant index evaluation retains its bounds checks. All 142 Flat-evaluator tests pass in surfaces-matrix-scope-candidate-1.log.

This remains candidate evidence. The initial lint checks exposed excessive nesting in checked dimension construction and an oversized AST display method; both were refactored without relaxing lint policy. Final lint/core, original Surfaces, fixed-canary, and complete-cohort checks remain pending at this checkpoint. No additional MultiBody model pass is claimed.

The final affected-package all-target/all-feature Clippy run passes (surfaces-matrix-clippy-3.log), as does the complete compiler-core rerun (surfaces-matrix-core-2.log: 484 passed). Formatting and whitespace checks pass. The normal-budget original Surfaces attempt is next; the matrix repair has no original-model or canary result yet.

The normal-budget multibody-array-constructor-surfaces attempt clears the constant binding failure and stops later in ToDae ED019: v1 has a colon axis without a call-site equality during function shape proof. Compilation takes 2.264 seconds; no simulation or comparator executes, so parity remains unmeasured. This is progress to the next rejected owner, not a model pass.

The fixed multibody-array-constructor-canary passes at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, with working-tree digest 3a5f684d12f4fb8dac00342ac6cc64214db6fc43b641c22960879e00a787826b. All twenty phase, simulation, and band outcomes match multibody-constant-slice-canary; all nine compared models and 175 trajectory channels remain high, with zero minor, deviating, missing, skipped, excluded, or nonidentifiable comparisons. The retained delta receipt is .git/multibody-campaign/surfaces-matrix-canary-delta.json. Tier 1 for the explicit constructor representation is complete. A full 566-model milestone sweep is next; the current complete-cohort count remains the earlier result.

Matrix-constructor cohort regression and reduced dimension proof

The subsequent target/msl/multibody-array-constructor-full sweep completed all 566 targets but failed its Flatten floor: 482 passed against the required 486. It ran at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, dirty-tree digest 22d63be38ba657693753172b6393649617582b65d83b38abfe274c0f65ce2a8d. The comparator ran: 143 models compared high, 16 tracked exclusions, zero missing or nonidentifiable traces, and zero deviating channels. All 566 model bands match the previous full run. MultiBody retains 14 high models with all 5,883 compared channels high, but OneAxis now fails Typecheck and FullRobot times out in Flatten. This is a failed regression gate, not release evidence. The complete phase delta is retained in .git/multibody-campaign/surfaces-matrix-full-status.json.

OneAxis and four Clocked drive examples share unresolved KinematicPTP2 nout dimensions. The reduced matrix_reduction_dimensions suite stops at the first divergent owner, Typecheck: five literal/local-scalar controls pass, while q_end={Constants.pi} fails ET004 when pi=2*asin(1.0) cannot be folded by the early evaluator. Requiring every array operand’s shape exposed a missing declared-scalar fact; the previous constructor had guessed scalar for unknown operands. OpenModelica accepts the exact literal and qualified-constant sources and produces the expected scalar q=time trace in both cases (matrix-reduction-omc/comparison.json). The initial test that proceeded into DAE instead reached a separate unsupported fill dimension and is not used as proof of this Typecheck regression.

The repair under validation supplies scalar-component facts by resolved DefId, checks that every qualified prefix is a class namespace, and retains array subscripts throughout the type-alias chain. Governing requirements are MLS §10.1/§10.4, SPEC_0007 Stage 2, SPEC_0001 identity domains, and SPEC_0033 §2/§6a. No floor, timeout, comparator tolerance, or target-list change is proposed.

The first declaration-based candidate passed the six original reductions and 230 evaluator/typechecker tests, including negative array-alias and component-array-prefix controls. Strengthening the function-result control to use asin(u) exposed the same missing fact for a scalar parameter: the core run had 489 passes and that single ET004 failure (matrix-reduction-core-1.log). The producer now records proven scalar component declarations independently of variability or foldable value. The strengthened six-test suite passes in matrix-reduction-dimensions-candidate-2.log. OMC also accepts its exact source and produces path.q[1]=time (matrix-reduction-function-omc/comparison.json). The strengthened candidate passes all 490 compiler-core tests (matrix-reduction-core-2.log) and all 230 evaluator/typechecker tests (matrix-reduction-libraries-3.log). All-target, all-feature Clippy passes after extracting the alias traversal and class collection into smaller helpers (matrix-reduction-clippy-2.log); the shape predicates are unchanged. Original MSL regression attempts, canary, and cohort validation remain pending.

The nine-model originating run multibody-declared-dimensions-origin restores OneAxis and all four Clocked drive examples to ToDae. It has no completed simulation or comparator measurement. Four regressions remain in that snapshot: PositionControlledDCPM reads a scalar record field (driveData.wMax), Dryden’s a constructor contains scalar powers (L^2/V^2), and SimpleLiquidWater / TestGlycol lose Medium.nX during Flatten.

Each remaining case now has a reduced producer failure. The record-prefix test returned unknown for a scalar field of a scalar record while retaining the array-record negative control (matrix-reduction-record-prefix-red-1.log). The AST dimension walk had no ordinary-power case at all (matrix-reduction-power-red-1.log). The reduced Medium package passed Typecheck and failed Flatten with medium.X: MatrixReductionDimensions.Medium.nX (matrix-reduction-medium-red-1.log, eight other controls passed).

The candidate now gives Typecheck and Flatten the same immutable rumoca-eval-ast::eval::DeclaredDimensions proof. It follows resolved type/base identities and preserves declaration/type-alias array subscripts; every parent reference must prove a namespace or scalar component. No constant’s numerical or string value is needed to establish its declared rank. Ordinary powers preserve a proven scalar or square-matrix shape only with a scalar exponent; vector bases and rectangular matrices remain unknown. These rules follow MLS §10.1/§10.6.7–9. The strengthened nine-test suite passes in matrix-reduction-shared-candidate-1.log; original-model and cohort validation of this final shared implementation remain pending.

The final shared implementation passes 922 library/integration tests (two existing ignored tests), all 493 compiler-core tests, and all-target/all-feature Clippy for AST evaluation, Typecheck, and Flatten. Logs are matrix-reduction-shared-libraries-1.log, matrix-reduction-shared-core-1.log, and matrix-reduction-shared-clippy-1.log. The exact record-prefix, scalar-power, and Medium-string fixtures all simulate successfully in OMC and produce the expected four scalar channels (matrix-reduction-extended-omc/comparison.json). Formatting and whitespace checks pass. No MSL count is promoted from these reduced tests.

The final same-list origin run multibody-declared-dimensions-origin-final restores eight of nine models to ToDae, including OneAxis, Dryden, and both Media models. PositionControlledDCPM still fails Typecheck because driveData is a replaceable record parameter; the current declaration proof deliberately does not treat replaceable components as known scalar. Its effective-rank obligation remains open. None of these nine models produces a completed simulation, so their parity remains unmeasured.

The fixed multibody-declared-dimensions-canary passes at HEAD bc71577f85df24957e5c9ab30fdaf4ed48da4311, dirty-tree digest f9e5eacffe7110ff80d51f704df928fd3cf4dec140af05db4104d54265ff0d4d. All twenty phase and band rows match multibody-array-constructor-canary. Nine models compare high and all 175 trajectory channels are high, with zero minor, deviating, missing, skipped, excluded, or nonidentifiable comparisons. The retained receipt is matrix-reduction-canary-delta.json. This completes Tier 1 for the shared declaration/power-shape repair; the remaining replaceable record case stays visible. The full cohort sweep is next.