AgentFEM Product Roadmap

AI-native repair loop (2026-11-12)

Preserve existing engineering semantics rather than create another model layer. The bounded sprint carries option and unsupported-capability diagnostics into execution records, retains units and engineering stages in inspection, and aligns ready-provider discovery with execution. Numeric export rejects invalid wrapped values; model publication is atomic and root-owned write failures propagate to MPI peers. Scoped capability queries and installed-callable signatures reduce unnecessary catalog reading and keyword guessing without a second API specification. Operational records share atomic publication; run preparation/finalization failures reach MPI peers, and final records must belong to the current run. The same publisher now owns Result manifests and checkpoint sidecars. Installed run/restart, corrupt-input rejection and VTK time-series checks close this bounded engineering slice; the sprint record retains both failures and acceptance evidence. Next measure agent task success and repair cost. Do not claim autonomy or token savings merely because structured error tests pass.

Fixed-horizon uncertainty and data gate (2026-11-08)

The bounded specimen workflow now handles fixed-plan zero-failure tests using exact two-sided binomial intervals. It shares record validation with survival analysis and rejects early censoring. Public P92 data search has not yet admitted a compatible replicate/run-out cohort; do not fit a scatter model just to fill that gap. See the source and acceptance note.

Specimen life evidence: bounded first slice (2026-11-07)

LifeObservation and life_survival now separate experimental failures and right-censored run-outs from simulated damage. The first slice provides exact condition/batch grouping, Kaplan–Meier estimates and pointwise exponential Greenwood confidence limits; see the workflow. Independent analytic fixtures verify censoring and intervals. The P92 DB adapter retains 18 source records in 18 conditions, not fictitious replicates.

Next: admit compatible replicate or censored experimental data, then evaluate one justified parametric/covariate model with held-out checks. Do not expand to general reliability software or claim calibrated P92 tail life from this slice.

P92 source, precision and runtime checkpoint (2026-11-07)

The original DB P92 source now supplies a coherent pipe cohort with rupture, fatigue and tensile-dwell observations; the fixed-parameter audit retains the published discrepancies rather than fitting them away. Generic creep increments now have small-increment precision evidence. The research material provider also removes duplicate local evaluations: complete bounded trajectories take about 34% less wall time on two tested meshes with unchanged accepted arrays. This is not a composite-contact runtime claim.

Next close the integrated software candidate, then resolve one specific life study: the DB cohort needs within-hold stress response and an applicable rupture-stress range before time-fraction coupling. Keep the observed nonuniform reaction convergence separate from unresolved clamp/relaxation spatial errors. Do not expand the material catalogue or fit the existing life labels again as a substitute for those measurements.

Data-admission refinement (2026-11-07)

Continue explicit cross-source research when same-heat P92 data are unavailable; retain pedigree, transfer assumptions and scatter rather than treating a preferred dataset as an absolute software gate. Do not convert that research into a component allowable. The evidence map records the industry-guidance basis and distinguishes numerical endpoint time from physical failure strain. No additional fit to existing life labels is planned.

P92 nonuniform structure checkpoint (2026-11-07)

The bounded bending screen records four mesh levels, time refinement and a small-model exact restart. Total ramp/hold reactions meet the predeclared numerical screens; the relaxation drop and clamp hotspot do not yet have spatial convergence evidence. Keep those limitations visible instead of expanding the mesh ladder without an engineering validation dataset. Same-point tensor histories now connect to creep assessment through the existing Result contract (PR #167).

P92 sequential-loading research checkpoint (2026-11-06)

The prior-damage dataset audit is merged (#165). The next bounded equation reference has independently checked time-domain and strain-domain integration, while retaining the source backstress inconsistency and a 3.67% as-received prediction discrepancy. Prior-state microstructure inputs remain explicit; the cited cyclic model was validated for P91 at 400–600 C, not this P92 cohort at 650 C. Resolve or bound that transfer before promoting a material law. Continue reusable software work without calling a numerically solved research reference an engineering life qualification.

P92 engineering admission checkpoint (2026-11-06)

The current P92 evidence map separates executable response and life baselines from engineering admission, and defines the minimum next data package. Generic tabulated rupture curves and the reviewed public rupture nodes are merged (#157/#159). Preserve their material-condition and stress- measure boundaries; do not combine them with a different fatigue cohort by name.

The published tensor reference, within-source life holdout and long-cycle response screen are integrated as research evidence. The next step is not another fit to the same labels: compatible independent rupture/fatigue/dwell data are the remaining engineering calibration gate. A newly reviewed open thesis supplies useful alternative cohorts but also inconsistent parameter tables; neither is silently promoted. See the source admission audit.

Generic small-strain point histories (#152), prescribed-temperature composition (#155), thermal clock guards (#160/#162), restart identity (#163), and structural cycle evidence (#161) are merged. These reuse one accepted clock, material transaction and result contract. The combined candidate passed targeted tests, isolated installed use, continuous/restarted result comparison and real VTK reading; see the integration acceptance record. This is source integration, not a new release or P92 service-life qualification.

Next gate	Work	Completion evidence
Engineering life	Select an application, compatible independent observations and a failure criterion before further calibration	Frozen calibration/holdout split, prediction errors and scatter, justified application-specific acceptance criteria
Component assessment	Use co-located accepted histories in a nonuniform component within the admitted material scope	Spatial/time refinement, boundary-condition checks and material/structure agreement
Further combinations	Prioritize temperature-dependent laws, regional materials, contact or MPI only against a concrete need	Separate capability contract, rollback/restart and numerical validation for each admitted combination

Do not repeat fits to the same eight life observations to fill the independent data gap, or expand all combinations at once. The current software foundation is ready for the next evidence-backed material study.

Research update: P92 tensor admission (2026-11-05)

The tensor admission record now connects the public scalar recovery equations to an explicitly assumed small-strain isotropic J2 extension. Local Jacobian/tangent checks, unit and basis invariance, scalar reduction, and a homogeneous structural hold/restart pass through the ordinary physical-time consumer. The implementation stays in research tools, not the material catalogue: Poisson ratio and multiaxial extension are assumptions, and experimental multiaxial validation, temperature interpolation and service-life prediction remain open.

Integration checkpoint: #147–#150 are merged. The tensor research candidate additionally passes structural time refinement toward an independent continuous reference at both published temperatures. Next prioritize additional fixed- parameter cyclic experimental comparisons and co-located small-strain histories; do not infer lifetime or multiaxial calibration from the homogeneous patch tests.

Direction

AgentFEM is building a readable, dependable, and extensible finite-element platform for people and AI agents. It does not replace the finite-element kernel already provided by FEniCSx, Basix, PETSc, and MPI. AgentFEM owns the engineering language, numerical procedure, state lifetime, result contract, verification evidence, and extension boundary around that kernel.

The public workflow remains:

Study -> Model -> scientific assets -> model.step(...) -> SimulationResult

Capability claims are promoted by executable evidence. A formula, an example, or a successful run is not by itself a validated engineering capability.

Foundation evidence log

The dated entries below preserve the development sequence. The admission checkpoint above defines current priorities; historical next steps are not additional release requirements.

Small-strain accepted histories (2026-11-06, integrated)

The physical-time small-strain user-material route now composes with the existing selected-point recorder: raw stress, total strain and named history share accepted time and atomic restart. See the design and evidence. This removes a data-flow gap for future life assessment; a material-specific rupture/fatigue law and independent validation are still separate requirements.

P92 long-cycle response screen (2026-11-06, integrated research)

Fixed published parameters were compared with sparse long-cycle experimental peak markers, reusing existing Radau/BDF campaigns. Reading sensitivity, solver agreement and nonzero response discrepancy remain visible in the cyclic screen record. Observed life normalizes the source figure; it is not predicted life. The next engineering gate is compatible, independently validated life data, not repeated tolerance tightening or refitting the validation markers.

P92 life-data gate (2026-11-06 research)

A fixed six-calibration/two-held-out baseline now evaluates the published combined creep-fatigue life table without reusing held-out labels. See the source, results and stop conditions. This is not an engineering material card. Prioritize independent compatible life evidence and same-point structural histories over further tuning of the same eight rows; do not add creep damage to an already combined life relation.

Physical-time small-strain user materials (2026-11-05, integrated)

The bounded consumer reuses the ordinary small-strain Newton Step and quadrature transaction. Explicit duration separates normalized loading from physical time; initialization, failed trials and restart do not advance the material clock. Exact accepted stress/tangent/state restoration uses a v2 same-layout serial checkpoint, leaving the rate-independent v1 route intact. Two local rate laws share this consumer; independent relaxation/creep responses, time refinement, cutback and corrupt-checkpoint tests are the promotion gates. The design record and examples/small_strain_relaxation retain the exact scope.

Selected-point histories and prescribed temperature are now integrated. Regional materials, contact and MPI remain separate composition gates. The P92 tensor reference has bounded homogeneous structural verification, but independent multiaxial experimental validation remains open. This source integration is not a P92 qualification, service-life prediction, or a release announcement.

Published P92 figure screen (2026-11-05)

The unchanged published-parameter reference now has a quantitative, traceable source-figure screen. Six authors' model-dash samples agree inside pixel-reading sensitivity; 24 first-cycle experimental markers across three held-out dwell conditions show about 5-6% sampled mean absolute relative error and local differences up to 33 MPa. These sparse samples are not a full-curve validation, a general material admission or a life model. Source coordinates, calibration, uncertainty scope, reproducible code and disagreement are retained in the figure-screen record.

PRs #142, #143, #144 and #146 are merged; #144 passed full FEniCSx and macOS installed-use CI. Co-located finite-strain histories are merged in PR #147. Next scientific gates remain full cyclic response and explicit multiaxial formulation/structural verification, not repeated refitting of a failed holdout.

Co-located finite-strain user-material histories (2026-11-05)

The next bounded consumer reuses results.material_point in ordinary finite-strain UserMaterial Steps. Raw S/P/F/E_GL, state and supplied uniform temperature share each accepted time and existing checkpoint/result contracts. Only selected points are recorded; sparse visualization does not decimate the history. No new solver or file format is introduced. Total Green–Lagrange strain is not automatically a mechanical strain-life input. See the design and evidence record.

Engineering PRs #142, #143, #144 and #146 are merged. The uniform-temperature composition passed full FEniCSx and macOS installed-use CI. Published recovery equations remain a numerical reference; quantitative experimental agreement and a verified multiaxial provider remain open scientific gates. Do not infer either from history output.

Prescribed thermal input for history materials (2026-11-05)

This bounded composition reuses the existing finite-strain material Step: temperature= accepts a positive kelvin scalar or a frozen, uniform linear physical-time history. Material updates receive endpoint temperature and its accepted-start increment; thermal knots, cutback, output and restart use the same accepted-state lifecycle. The four-file example exercises either history law with an optional thermal cycle, without a project-level solver.

Local analytical expansion/reaction, plane-strain/plane-stress, tangent, rollback, restart and standard project tests pass. This is prescribed thermal input, not a heat solver or experimental material validation. Spatial thermal fields, regional providers, contact and MPI compositions remain separate gates. See the implementation record.

Candidate mechanism screening and solver robustness (2026-11-05)

Engineering slices #138–#141 are merged. The separate Windows-guidance PR #142 has passed its scoped CI; it is not a new clean-host installation trial. The P92 sinh screen tests existing flow laws before introducing new material mechanisms. Four RF-only candidate fits still miss the long low-stress SLR hold; SLR remains diagnostic rather than blind validation. Full-path FEM time refinement now passes after a general creep Newton correction: impose the new prescribed displacement before line-search evaluations, while retaining accepted-state rollback.

Next: discriminate joint plasticity versus physical-time recovery; specify and verify any genuinely missing mechanism; then review prescribed temperature histories for user materials as a separately tested composition. Keep numerical verification, experimental material validation and life assessment distinct.

The static-recovery discriminator now isolates strength decay from stress relaxation: a strain hold relaxes only after yield reactivation in this bounded tensile model. Analytical activation, time refinement and local tangent checks pass. Keep the tool outside the core; next audit the complete public P92 recovery model and its own experimental conditions before implementing or fitting a material provider.

Co-located histories and real-material gate (2026-11-01)

Keep the preceding output/integration PR separate from this scientific slice. The new opt-in material_points contract records selected raw integration-point stress, total strain, internal variables and prescribed kelvin temperature at accepted physical times. It reuses ordinary result histories and both existing creep checkpoint formats; corrupted history or changed selection rejects the whole restore. Serial thermal-history restart now compares the complete path identity, not the receiver's current temperature sample. See the history contract.

The first composition tests nonuniform time-dependent temperature with regional pure creep, or homogeneous joint creep/plasticity. It does not broaden the joint material to regional, contact or MPI use, nor admit temperature fields on arbitrary finite-strain UserMaterial Steps. Sampling is selected-point only, every accepted increment, independently of visualization frames.

The Alloy 617 source audit selects a public report with test conditions and model parameters, but identifies softening/damage mechanisms incompatible with a direct parameter substitution into the current hardening-only model. Real-data calibration and held-out validation remain open, not replaced by synthetic test passes. A separate public P92 experiment now executes a checksum-pinned RF fit and SLR holdout comparison through an effective relaxation baseline, with independent FEM time refinement and restart. This is conditional hold response, not a full-cycle or lifetime validation. Next: review elastic and softening behavior, freeze a compatible local law and a new holdout, then validate the complete loading path through the same structural lifecycle.

The full first-cycle diagnostic now executes RF and SLR without hold-start stress resets, using ordinary creep Steps and an independent continuous scalar reference. Three-level time refinement passes, but the high-stress Norton fit does not transfer to the long low-stress SLR hold. Measured-path and modulus sensitivity checks do not remove that discrepancy. Preserve this as a negative transfer benchmark; screen existing flow/plastic assets before adding recovery or softening, and freeze a new validation condition before retuning. Engineering integration and scientific material promotion remain separate: #137 and #138 are merged. This work does not claim a validated P92 material, lifetime prediction or a new release.

Material-only structural workflow sprint (2026-10-29)

User feedback identified an important boundary: a local history law should not force each project to implement Newton, quadrature transactions and XDMF layout. The bounded workflow reuses the existing finite-strain UserMaterial protocol, shared Newton/PETSc integration and committed/trial quadrature storage. It adds no hydrogel-specific model family and does not infer multiaxial physics from a one-dimensional fit.

Priority	Deliverable	Admission boundary
P0	One ordinary model.step workflow for two replaceable history laws; positive physical duration, cutback and same-layout restart	Experimental serial homogeneous 3D/plane strain, strong BCs/reference dead loads; fixed-old-state dP/dF is mandatory
P0	Single-grid accepted multi-field output, readable closed frame files, explicit recovery and restart provenance	Real VTK/Warp and interruption checks; no fabricated unsaved fields, energy, or independent validation
P0	Extension-aware upgrade guidance and a four-file runnable example	Recognize explicit Step lowerers, not entire extension directories; local materials do not need project Step providers
P1	Bounded condensed plane-stress adaptation, in the same material-only workflow	Explicit reflection-symmetric 3D law; positive transverse stiffness; scalar thickness equilibrium and full-update Schur tangent; independent local/structural, rollback/restart and real VTK checks
Next	Extend independent scientific coverage before wider admission	Synthetic nonuniform damage space and viscous time convergence are covered; next validate selected real provider laws with their owners. No arbitrary anisotropy, shells, hidden 1D-to-3D inference or batch speedup claim
Later	Temperature/external-field histories, regional materials and distributed global equilibrium	Add each composition with atomic failure/restart and independent scientific evidence, not by relaxing admission checks

This sprint complements thermal/cyclic-creep work below; it does not replace its physical-time semantics or introduce another state/output framework. It does not close the outstanding nonuniform plastic-contact spatial gate or promise a composite performance improvement. Precise evidence is recorded in the sprint record. The bounded reduction's design and evidence are recorded in the plane-stress record. The example also covers ordinary CLI pause/restart/whole-run comparison with explicit partial-result trust, rather than requiring a project-managed output loop. Accepted-state ownership guards and unit-robust tangent evidence are engineering hardening, not additional material validation or wider admission. Automatic accepted-state persistence now reuses checkpointing.every through the same Step option contract. The four-file example retains two scheduled archives; failure, pause and restart do not need project checkpoint callbacks. Cadence is based on the restored accepted ledger; only newly published scheduled records are eligible for retention. Manual and previous-run files stay outside that policy. Disk/reporting failure is not numerical cutback or material failure. The follow-up verification audit also rejects observable provider input mutation without modifying the caller's request. Full/reduced finite-strain and small-strain tangent evidence now share unit-robust error arithmetic; scalar/batch stress symmetry retains point-local scales. This is verification infrastructure, not experimental validation of any user law. Reusable local rotation evidence now requires explicit invariant/spatial-vector/spatial-tensor declarations for every history variable and rejects undeclared external fields. It checks each channel independently, including two different endpoint rotations; it does not replace derivative, thermodynamic, path-wide or experimental checks. The existing J2 material is included as a local cross-check, not a wider structural admission. Real material requests/responses and checkpoint restoration now refuse complex values before conversion, instead of silently discarding imaginary components. This is boundary validation, not support for frequency-domain constitutive laws.

The next output-control slice adds output_every to the same Step contract, with no project callbacks: accepted-increment sampling, segment endpoints and continuation boundaries are separate from physical-time discretization and checkpoint retention. Default every-increment output remains unchanged. This reduces recovery/storage work when fewer frames are requested; it is not a constitutive-solver or composite-speed claim. Admission remains experimental and serial. See the output-control record for verification and measured scope. The same slice removes repeated serial geometry-expression preparation in QuadratureField, while evaluating current geometry every time. It does not cache physical weights or change MPI collective preparation. Same-frame short- path confirmation measures about 13.5–15.7% less total time for the two educational laws; this is not a composite full-run or portable performance promise.

The integration checkout uses 0.4.5.dev0, not the already published 0.4.4 identity. Its development contract retains every existing admission limit and records the new workflows as experimental. This is not a release or promotion. The integration gate runs the ordinary four-file example from an installed candidate with no source-path injection: two laws, planned pause/restart, continuous and sparse output, accepted-state comparison and VTK/Warp reading. Reference geometry remains fixed after Step construction; changed geometry and lossy endpoint conversion are refused before state or output mutation. See the integration audit for installed-use evidence and the distinction between passing gates and pending integration.

Thermal–inelastic continuity sprint (2026-10-29)

PR #132 maintenance is merged at 44fda9c4; the published 0.4.4 tag remains unchanged. The next bounded sequence is reliability → thermal/cyclic-creep integration → damage/life assessment. It does not reopen packaging, claim commercial contact parity, or replace the earlier open spatial-accuracy gates.

Stage	Implemented slice	Remaining promotion gate
Reliable results and restart	Validated zero-copy XDMF continuation views; atomic late-failure creep checkpoint restoration	Installed-use regression; a joined view alone does not prove physical restart continuity
Thermal–cyclic plasticity–creep	Shared implicit Sinh flow; simultaneous J2/Chaboche + creep update, consistent tangent, ordinary model.step, temperature history, cutback, separate energy channels and serial restart; analytical time checks and a nonuniform active-plastic/creep spatial reference	Independent material calibration and nonproportional thermal spatial references; MPI, contact, regional materials and temperature-dependent plastic properties remain unsupported for the new composition
Damage/life	Bounded scalar strain-life and time-resolved creep-damage assessments reuse rainflow and the existing interaction envelope; explicit units, source sampling and domain rejection	Paired mean-stress/multiaxial/nonisothermal life and global damage feedback need their own contracts and independent validation; no material calibration is implied

The initial joint-material gate is a homogeneous 3D small-strain hold/reverse fixture, including an independent uniaxial stress solution. It is not an industrial lifetime prediction, two-way heat generation, or a performance benchmark. The design and evidence record tracks exact scope. The ordinary examples/thermal_cyclic_creep project uses synthetic parameters, not confidential calibration data.

The creep/fatigue verification ladder now ties the next gate to independent Norton relaxation, cyclic loop work and time-fraction references, followed by compatible held-out material test data. ASTM catalogue references identify test families, not software certification; the generic linear interaction is not a qualified design-code envelope. The first open rupture-data audit found missing test metadata and a material label discrepancy; it is recorded as a candidate, not admitted as a validated alloy. Use independently documented deformation, cyclic and dwell tests to decide whether static recovery or cyclic softening must be added next; do not silently fit those mechanisms with the currently hardening-only plastic asset. The prescribed nonuniform-temperature joint field now also has an independent spatial-refinement reference. It isolates Arrhenius rate dependence with zero thermal expansion, not a full thermal-cyclic component qualification. Future material comparison must distinguish ordinary strain-controlled tests from elastic-follow-up/SMT tests; the ANL-ART-168 reference provides that distinction. Before nonuniform component-life productization, preserve co-located accepted stress/strain/temperature histories with point identity through restart; separate spatial maxima are not a replacement. Reuse result/state sampling contracts, not a second simulation or sensor framework. An independent rotating-stress material reference now also checks nonproportional J2/Chaboche plus creep under heating/cooling, including distinct plastic/creep directions and first-order time convergence. This strengthens local integration evidence; the nonuniform structural/experimental promotion gate remains open. The existing regional pure-creep route now has an independent elastic-follow-up series-shear reference. This verifies structural restraint effects without adding a case-specific correction factor or admitting regional joint plasticity.

0.4.4 published; bounded maintenance (2026-10-27)

0.4.4 is published at immutable commit 7dbf936df18997bd8785a72f27ff230bcad68581. The publication workflow completed successfully. The release scope continues to separate admitted experimental compositions from stable engineering changes. Follow-up maintenance reviews the changed assembly, state and output paths; it does not replace the published tag or reopen installer/package development. Do not start another large spatial sweep or a full composite run merely to close a maintenance review.

The graded-cycle path study is complete within its recorded scope. Spatial plastic-contact accuracy and an independent same-input nonlinear reference remain open promotion gates; they must not be relabeled as passed. Correctness regressions in admitted routes still block release. Fresh Windows/WorkBuddy field acceptance remains separate from installer-contract CI. The current historical release board and evidence are in the release decision record. Earlier dated boards below are historical evidence, not simultaneous tasks.

The short post-release review found a bounded assembly-validation gap: individually finite duplicate contact stiffness entries can overflow when summed, and complex inputs could lose their imaginary part during conversion. Reject both before PETSc mutation; retain the existing batched insertion and test joint rollback/continuation. This is reliability hardening, not a new speedup or promotion of experimental contact accuracy.

Windows onboarding simplification (2026-10-26)

The new user field report identifies Ubuntu distribution metadata download, not the numerical stack, as the main installation blocker. Promote the already reported TUNA image/import route to the short onboarding path: WSL platform → checked Ubuntu image → ordinary account → mirrored environment and installed first solve. No checkout, optional helper or Windows Features checklist is a prerequisite. Retain mirror choice in local handoff notes and offer a one-action manual fallback only when the agent lacks permission. Detailed installer recovery remains available separately. The report's root installation and older package success are evidence, not recommended defaults or proof of current-release/project compatibility. New non-root quickstart needs clean Windows/WorkBuddy acceptance; local documentation checks and PowerShell parser CI do not substitute for it. This work does not change solver algorithms, contact maturity or simulation performance claims.

PR #125 is merged at 5e6cba5b after all applicable checks passed for 11accb3b, including PowerShell 5.1/7, Linux/FEniCSx and macOS installed-wheel acceptance. This is engineering integration, not a new Windows host field trial.

Active sprint decision board (2026-10-26)

The numerical follow-up starts from merged main a862e123 (including PR #123). It prioritizes bounded diagnostic and execution improvements, not another capability expansion. CI establishes source/installation regressions, not nonuniform spatial accuracy or commercial parity.

Order	Work	Decision gate
P0	Keep engineering integration separate from scientific maturity	The release contract names periodic J2/contact as experimental; no general contact or mesh-independence promotion
P0	Separate path-work error from spatial contact error	Independent elastic assembly/MPC audit, per-increment work/energy accounting, then one graded plastic path refinement at common stations
P1	Remove redundant observation work	Scalar energy/feature histories must not assemble unused global tangents or copy whole fields; retain current geometry, accepted-state transactions and legacy observers
P1	Preserve ordinary-use acceptance	Reuse unaffected installed output/restart evidence explicitly; validate changed assembly paths without AI-provided executable glue
Exit	Commit tested improvements and retain open science gates	No release tag or commercial-parity claim on partial validation; reuse unaffected full composite evidence

Current progress: the independent discrete elastic audit agrees with production force/energy on uniform and graded 8/12-cell meshes. A primal affine J2 energy ledger now distinguishes stored energy, plastic/friction loss, penalty energy and accepted external work. Portable two-phase 1↔2-rank checkpoint checks retain its complete scalar history; this does not admit distributed implicit contact. Scalar contact-energy observation no longer needs full residual/tangent assembly.

The initial-yield diagnostic explains almost all first-station work imbalance before contact in the tested graded fixture. A same-source small full-cycle control reduces the initial increment alone, with 232 → 237 accepted increments: final relative energy residual falls from 0.2106% to 0.0298%, while final force changes about 0.0014%. This is an accuracy diagnostic, not a runtime-speed claim or justification for a new universal default. The graded full cycle now also completes: joint initial-step and plastic-cap refinement lowers the final relative work/energy residual from 0.2046% to 0.00625%, with a 0.1137% final-force change. This separates path accounting error from spatial sensitivity; it does not close the spatial or independent same-input reference gate. Details, raw evidence and exact limits are maintained in the contact sprint ledger.

The observation optimization has a separate 64-increment ABBA confirmation: 4.73% less solve/observe time within the energy-enabled candidate. A net comparison against merged main measures 3.68% less time, including the new energy history, with identical physical arrays and common accepted histories. These are bounded serial implicit-fixture results, not additive percentages or explicit-composite speed claims. The final merged-main candidate also passes 163 installed regressions and the ordinary init → check → run workflow; CLI discovery no longer initializes MPI merely to display help or version.

Windows onboarding is a separate bounded engineering change, merged in PR #125: reuse WSL/Ubuntu, honor an explicitly chosen mirror, remove the mandatory source checkout, and give one manual action when permission/download blocks a stage. A fresh Windows/WorkBuddy machine remains a user acceptance check; PowerShell contract CI alone is not such a test.

The fixed-width edge-belt diagnostic must also record its sampled area: a discontinuous indicator has mesh-dependent quadrature error. Complement it with a continuous polynomial edge/center partition of unity whose reference areas are known analytically. This diagnoses pressure redistribution without inventing an error estimator or changing the contact law.

Historical entries below remain evidence, not a second current backlog.

Current integration results: SVD reuse passed independent history/rotation/ conditioning and global-path checks, with 3.15% less solve-and-observe time in the fresh 64-increment ABBA confirmation. It changes neither the material law nor the accepted path. A native wheel from 51c4a911 was installed into an isolated AgentFEM environment with shared numerical dependencies. Ordinary agentfem run partial/restart/continuous paths agree in displacement, material state, contact force and work history. Installed output, interrupted publication, history/retention and independent contact-reference suites pass. This is not a clean Windows installation or new full-composite performance proof. Details and the reconstruction protocol are in evidence/contact/2026-10-25-installed-periodic-contact.json.

PR #118 is merged at 897f1f74 after final source d8392e98 passed all applicable checks in run 40121006505, including Linux native distribution, FEniCSx and macOS installed-wheel acceptance. The final serial suite reports 4988 passed and 428 skipped; optional-native and rank-specific skips are not native coverage. Separate native and distributed gates pass. This is engineering integration, not a 0.4.4 release tag or promotion of experimental contact. An intermediate Linux run exposed one ill-conditioned first-Piola component comparison failure (4961 other tests passed). The failure is preserved; an independent high-precision diagnosis and per-point objective tensor comparison replace that componentwise test, without changing the runtime or its tolerances. Fresh Linux confirmation at e4207c56 passed 4976 serial tests and subsequent distributed/release-smoke checks in run 40115587874. The final-head confirmation above also covers the later boundary-admission safety changes. See the precision follow-up in the sprint ledger. The workflow and material guide also separate finite-strain contact from small-strain J2/Chaboche and defer geometry/friction/MPC combinations to one current capability matrix rather than stale one-tool restrictions.

The release-scope audit also found a correctness hazard: several small-strain material providers advertised boundary Operators they did not assemble. Native J2/Chaboche, power-law creep, quasistatic/harmonic Maxwell and neutral small-strain providers now reject these combinations before Step construction, with one addressable reason shared by capability inspection and compatibility builders. This is a necessary safety correction, not a new contact capability.

The spatial gate is narrower, not closed: the controlled 4/8-cell plastic first-station force gap is 6.52%, with different sampled contact activity. On the same coarse mesh, supported contact quadrature 7 -> 11 changes force by only 0.00557%; halving the coarse plastic increment cap changes force by 0.02025%. Raising coarse bulk quadrature 6 -> 8 changes force by only 0.0008892%. None explains the spatial gap on its own. Independent polynomial P2/Q2 trace tests pass force, moment, virtual-work and tangent integrals. The bounded 12-cell continuation now completes the first station with 305 accepted increments. Forces on 4/8/12 meshes are 4.48528146 / 4.19262706 / 4.10251562 N; successive coarse-relative differences decrease from 6.52% to 2.15%. The 8/12 sampled active areas agree, but peak pressure continues to grow and edge-weighted redistribution exceeds the total force change. This is a useful trend, not a continuum-error bound or full-cycle accuracy certificate. Evidence: evidence/contact/2026-10-25-periodic-plastic-pressure-moments12.json.

Next bounded science gate: stop uniform refinement here. First compare contact-edge resolution against the same physical patch and controls, using integrated reaction and smooth pressure moments rather than peak pressure alone; then obtain an independent same-input reference. Any claimed improvement needs its own fine-path sensitivity check. Do not simultaneously open a new contact law, penalty sweep, full-cycle sweep or larger uniform mesh. Promotion remains separate from the engineering fixes already merged.

The diagnostic now supports an edge-graded mesh with unchanged cell count, physical patch, periodic trace equations and material/solver controls. It is research fixture geometry, not a new public mesh generator. Positive cell quality, physical volume/patch areas and checkpoint mesh identity have focused tests. Elastic 8/12-cell force differences decrease from 1.01% (uniform) to 0.60% (graded); this is a refinement comparison, not a continuum-error bound. The graded plastic first station completes through an ordinary checkpoint continuation. With 14,739 displacement DOFs its reaction differs from the 46,875-DOF uniform-12 comparison by 0.416%, versus 2.196% for uniform 8. The finer uniform result is not independent truth; sampled peak pressure continues to rise. An independent same-input reference and graded-path increment sensitivity remain open. No larger uniform sweep is started. Full details and the interrupted/continuation evidence are in the contact sprint ledger.

The contact execution candidate sums current quadrature matrices sharing an ordered nodal stencil before generating scalar COO entries, then inserts the coalesced rows through PETSc's subset-CSR interface. It changes neither contact quadrature, projection, physical law nor state ownership, and introduces no new user knob or dependency. A 32-increment cyclic screen reduces submitted COO entries from 7,661,700 to 340,200 and solve/observe time by 5.73%; independent 64-increment confirmation measures 5.48% less time with the same 208 Newton corrections and matching four-station fields, states, contact force and penalty potential. See evidence/performance/2026-10-25-contact-stencil-confirmation.json. Full-matrix storage reuse alone is withdrawn: its small screen showed no meaningful gain. Do not add these percentages to earlier optimizations or transfer implicit timing to the explicit composite.

The existing J2 tangent backend now packs its nine derivative directions into small rectangular products and reuses their common transform, without another public tuning switch. Fresh serial cyclic P2/J2/plane-penalty/MPC ABBA evidence measures 8.32% less solve-and-observe time with the same 32 accepted increments, 113 Newton corrections and four-station field/state/contact comparisons. The 4096-point tangent kernel improves by about 35.78%, not the whole solver. The original near-zero-component comparison failure and its justified per-point tensor-scale follow-up are both retained; see evidence/performance/2026-10-24-j2-tangent-block-confirmation.json and the contact sprint ledger. Do not add this result to prior percentages or transfer it to the explicit-composite run. The subsequent SVD-reconstruction acceptance is recorded above and is a separate measured improvement, not an added percentage.

The user's 2026-10-23 clarification keeps solver capability and throughput first. A separate contributor is handling WSL installation: inspect that work for later integration instead of starting a competing installation sprint. PR #119 and its dependent #120 are now merged into main at d87f1c0e and a7f69a8a. Same-version build identity, installation identity, cross-package locking, bounded host probes and preserved restart paths have source-level regressions. The final onboarding head af36afde passed Windows PowerShell 5.1/7, numerical and installed-wheel CI. WSL/elevation remain mocked in those Windows tests: real clean Windows/WorkBuddy acceptance is still explicitly pending, and no new installer bundle is released. This closes source integration, not the customer-acceptance gate or solver sprint. The joined body/viscosity policy now has a complete late-state ABBA comparison (5.69% less mean advancement time, identical accepted fields and physical history). This is not a full-run gain. Current smooth-interior Q1 sign evaluation with two-ray fallback also preserves the complete late accepted state. Combined ABBA screening shows 4.83% less mean advancement time, not the sum of component improvements. Process-local worker screening did not select a faster budget; a versioned contact restart contract separates explicitly partition-invariant execution budgets from physical identity. Legacy checkpoints remain strict until resaved through their original runtime. No silent identity bypass is permitted. Keep the graded finite-J2 spatial study separate from both timing experiments. The immutable-array validation cache has been withdrawn: its isolated late-path comparison cannot distinguish the small mean difference from timing drift. Keep the snapshot-before-validation correctness fix without the cache. A Procedure-owned linear workspace now keeps current PtAP/tangents across frozen normal-augmentation equilibria and releases resources at attempt exit. Its small allocation-only gain enables a separate opt-in adaptive GMRES preconditioner policy. The isolated 12-cell/32-increment screen reduces mean time by 13.39%, with accepted station arrays within declared tolerances; fresh 48-increment confirmation of the integrated public route measures 13.73%, with 48 accepted increments and 533 Newton corrections in every run. This is an implicit-solver result, not an explicit-composite or full-run speed claim. The graded 16/256 study is now complete: contact tolerances pass, but final reaction changes 7.2893% from graded 12/256. Grading alone has not closed the spatial gate. Stop blind refinement/penalty tuning; the next discretization study must address compatible volume, contact-trace and periodic spaces. Do not silently use P2 volume interpolation with the current P1 nodal AL contract. The commercial-parity and independent-reference gates stay open.

The combined serial implicit execution route has also completed an isolated 12-cell/48-increment ABBA comparison: 24.32% less mean solve-and-observe time relative to fresh material trials, separate outer workspaces and direct solves. All four paths retain 48 accepted increments and 533 Newton corrections, and all four-station field/state/contact/work comparisons pass. This evidence supports the composition, not addition of its individual percentages, and does not change the explicit-composite timing claim. See evidence/performance/2026-10-23-combined-affine-execution.json.

The eight-cell P2 diagnostic now separates contact and bulk integration: contact quadrature 7 -> 11 changes final reaction by 0.0351%, then bulk quadrature 4 -> 6 by 0.00887%. These do not explain the much larger spatial gap, and defaults remain unchanged. A separate native worker-local nodal-gather prototype preserves the late accepted path but its 1.41% mean timing change is smaller than the 2.17% baseline drift; it is withdrawn rather than added as another public performance switch.

Further diagnostics distinguish two remaining errors rather than attributing everything to contact: at fixed eight-cell P2 discretization, 32 -> 64 -> 128 increments change the final plastic reaction by 6.41% then 3.16%. With plasticity disabled by a high yield stress, common Gaussian-penalty 8 -> 12-cell controls change reaction by 4.46% for P1 and 1.02% for P2. This supports investigating compatible volume/trace approximation and path refinement separately; it does not certify continuum accuracy, prove locking or authorize P2 nodal AL.

An output audit found a separate correctness issue: integration-point cell recovery normalized reference-rule weights without the geometric Jacobian. The physical-measure correction now passes analytical distorted Q1/Hex8 and curved-triangle tests, scalar/vector/tensor recovery, and two-rank checks for owned/ghost publication, an empty rank and collective invalid-measure rejection. Raw constitutive states and solver updates are unchanged; metadata distinguishes mesh-geometry weighting from any inferred deformation-volume weighting.

The next work-reduction mechanism is now an opt-in core increment policy, steps.automatic(inelastic_prediction_safety=...), using the last accepted plastic increment to cap the ordinary Newton-based next proposal. It does not relax the declared plastic-increment limit or change mandatory load knots. The bounded serial global-J2 strong/affine routes share explicit admission, accepted-state rollback and versioned restart identity. Regional/mixed/MPI prediction remains rejected, without restricting their existing policies. On a cyclic P2/J2/plane-penalty/MPC fixture, source-frozen ABBA means are 186.400 s ordinary and 114.687 s predictive (38.47% less), with rejected attempts 140 -> 13 and Newton corrections 1300 -> 806. Common-station reaction differs by at most 0.204% from a finite 512-increment reference; that reference does not establish continuum accuracy. This is a new increment algorithm, not another unchanged-path execution gain, and cannot be added to 24.32% or applied to the explicit composite runtime. Compact evidence is in evidence/performance/2026-10-23-inelastic-increment-prediction.json. Prospective 1024-step-reference confirmation retains the same predictive 232-increment path. Halving the plastic limit gives 449 accepted increments and reduces maximum reaction/displacement/mean-PEEQ differences to that reference from 0.202%/0.0384%/0.0291% to 0.104%/0.0119%/0.00166%. Fixed-reference 512 -> 1024 final reaction still changes by 0.405%: preserve the remaining path-accuracy uncertainty rather than selecting the more flattering reference. The next scientific gate remains compatible spatial approximation at controlled path resolution, not another controller knob.

The eight-cell P2 controlled-path follow-up reached common factors 0.25 and 0.5, then stopped at its 1800-second runner limit. At the first station, path refinement changes reaction by about 0.27% from fixed 128 increments, while the four/eight-cell controlled paths still differ by about 6.52%. Unloading opens contact correctly at 0.5. Preserve these partial observations and the timeout; they do not close full-cycle accuracy or spatial convergence. See evidence/contact/2026-10-24-periodic-p2-controlled-path8.json. No larger blind mesh/parameter sweep follows from this result.

Order	Work	Current decision / acceptance boundary
1	Retain unchanged material trials and scoped linear work	Implemented for bounded serial J2/AL; combined 24.32% timing reduction with station parity and mutation/rollback/restart guards. Unsupported materials/mixed/MPI use ordinary refresh; no scope expansion inferred.
2	Isolate spatial/path error and reduce rejected work	Spatial gate remains open. Accepted-plastic prediction now reduces rejected work on a bounded cyclic fixture; it is not an error estimator. Hold the physical patch fixed and keep volume/trace approximation and path-refinement evidence separate.
3	Explicit body/contact critical-path reduction	Selected joined-viscosity/interior-sign route remains a late-path result, not a new full-run claim. Reuse the 25-frame baseline; the new nodal-gather prototype is withdrawn rather than added as an unproven switch.
4	Bounded release integration	Solver/output and main-integration head 197ba2c8 passes its full remote CI, including numerical and installed-wheel gates. Subsequent tangent changes require their own remote checks. Scientific claims remain limited to passed geometry/law combinations; PR #118 remains a development candidate while spatial/reference gates are open.

The output/recovery review reproduced and fixed two additional failure paths: a closed serial writer could reopen and truncate its previous HDF5 file, and completion cleanup could prune earlier checkpoints before result publication succeeded. Closed writers now reject re-entry and writes; failed closes remain explicit, and repeated cleanup cannot promote an interrupted index. Completion retention publishes only surviving checkpoint references before deleting owned older files. Result-construction/publication failures preserve those files. The affected output/archive/project set passes 124 tests (one MPI-only skip). After interruption, the focused material reuse, linear workspace, increment prediction and checkpoint review passes 116 tests (one MPI-only skip). These are overlapping risk-based regressions, not new timing or scientific-validation claims. Existing generated site/ changes remain untouched.

The first retained-material source candidate is scoped to a single equilibrium attempt and compares evaluated old/new gradients, committed/trial state, current stress/tangent/energy buffers, material parameters and time. State retains no second response array, and the transient guard is never checkpointed. The contact residual and global Jacobian remain current; this is not Jacobian lagging. See tests/test_retained_material_trial.py and tools/benchmark_retained_material_trial.py for mutation and ABBA checks.

2026-10-22 implementation priorities

Keep the implicit-contact scientific gate and the explicit-composite throughput gate separate. A faster explicit implementation does not close implicit contact convergence; a completed composite run does not establish external-code parity.

Sprint evidence now includes a complete frozen 100,000-increment run in 3 h 24 min 56 s, with all 25 frames, accepted histories and checkpoint arrays equal to the preceding baseline. Independent VTK reads every frame correctly. The controlled late-path comparison supports 12.11% less advancement time; do not attribute the entire larger whole-run difference to code changes. Reusable changes cover sufficient cell admission, exact topology/coordinate reuse, operational restart identity, linear-option transport, prepared-operator failure cleanup and opt-in lossless history archives with a data-only reader. The new independent confined cyclic plastic/contact reference passes, while the inhomogeneous spatial convergence and external-comparator gates stay open. The complete 16-cell/256-increment contact path meets the original nonlinear acceptance bounds, but the 12→16 final reaction difference is still 5.2481%. Mean PEEQ and displacement L2 change much less (0.3729% / 0.2490%); neither is a substitute for reaction convergence. Preserve this unresolved distinction when choosing the next discretization/contact study. An explicitly separate elastic-limit cycle also retains a 6.5122% 8→12 reaction gap with zero plastic strain. Prioritize spatial discretization and the contact constraint representation; do not assume the J2 update is its cause. Exact active-normal elimination on those same elastic meshes agrees with augmentation to below 3e-8 N in total reaction, while preserving the spatial gap. Further penalty tuning is not the priority for that observed difference.

1. Measure and remove redundant execution. The completed 100,000-increment composite run still attributes approximately 89% of solver wall time to residual assembly. Profile the late damaged configuration, including the overlapping body/contact branches, before changing algorithms. First test whether a coherent body/contact execution change removes remaining work. The sufficient native whole-cell screen is now installed and verified; retain uncertain-cell fallback, the material law, increment, numerical mass and accepted-state contracts rather than repeatedly retesting this same gain.
2. Make execution evidence explainable. Distinguish inclusive body cost, geometry cost, owner-thread waiting and publication cost. These are concurrent or nested quantities, not additive percentages. Use the existing performance ledger; do not create a second authoritative solver/state or benchmark policy.
3. Close the actual implicit-contact accuracy gap. Continue the bounded finite-area-supported J2/plane fixture with common accepted load stations and independent spatial/increment refinement. Diagnose contact integration and material path errors separately before extending geometry or enforcement. Keep the independent comparator and unsupported sharp-feature boundaries explicit; do not promote the 0.4.4 parity target by test count.
4. Preserve recovery and usable output. Keep the verified 25-frame run and its final checkpoint unchanged. New profiling runs write to separate locations. A restart identity mismatch must be investigated, never bypassed to make a timing experiment work. The selected combined change has now passed frozen installed whole-path acceptance; another full run is deferred until a new measured change merits it. History encoding is an output choice, not a solver/physical policy or a substitute for saved field frames.

The execution decomposition follows the distinction between element restriction, quadrature work and global reduction in libCEED's interface concepts. Displacement-bounded candidate reuse is informed by LAMMPS neighbor-list maintenance, but any Q1 contact reuse requires its own complete-coverage and moving-geometry tests. Neither source supplies an AgentFEM speedup number.

0.4.4 bounded contact-parity sprint (target, not a capability claim)

The next release target is local, explicitly bounded parity for serial three-dimensional solid-versus-prescribed-rigid, quasistatic implicit contact with finite-strain J2. This means a user can state and independently verify the same supported physical problem through the ordinary Model/Step/Result workflow; it does not mean all Abaqus/Standard general contact, identical floating-point fields, or automatic admission of every triangulated tool.

The ordered nine-gate task and technical bibliography are maintained in the 0.4.4 contact sprint record. The October 26 independent elastic assembly audit now passes uniform and graded 8/12-cell controls, including periodic virtual work. This narrows the spatial investigation but does not certify the plastic full cycle. The next bounded checks remain graded plastic increment sensitivity and loading/unloading histories. The primal affine J2 result now records restartable accepted energy components and combines existing affine/tool work, with explicit unavailable status for incomplete historical evidence; this is diagnostics, not automatic scientific promotion. Promote only combinations that pass independent same-input comparison, mesh/increment/enforcement convergence, residual--tangent and lifecycle checks. A failed nonconvex sharp-feature gate narrows the release statement; it must not be hidden by a visually plausible contour or another penalty choice. A moving feature boundary in a surface integral and actual physical edge/vertex contact are separate numerical questions: passing one does not license the other. Distributed implicit contact, deformable--deformable/self contact, shell-thickness contact and arbitrary CAD remain outside this bounded target.

The current serial affine Newton implementation reuses the total residual and Operator responses only within an unchanged trial, avoiding a second contact evaluation before the tangent solve. Small frozen-old-Procedure comparisons cover fixed/moving sphere contact, friction and automatic plastic cutback; this is not an explicit-composite speed claim. A separate bounded analytical plane diagnostic now includes the actual MPC affine offset, distinguishing immovable penetration from mere multiplier redundancy. Neither improvement admits AL under arbitrary MPC; compatible multiplier spaces and their independent pressure/convergence evidence remain required. An option-contract audit also corrected explicit basic/no-search Newton: finite valid full steps may temporarily increase the residual, while invalid material trials still reject. Backtracking remains the default; serial and distributed affine paths share the meaning with PETSc, and affected legacy basic-mode checkpoints cannot silently change algorithm. The bounded periodic-plane route now has joint work/restart and explicit mesh/increment/penalty sensitivity records. A reaction audit distinguishes its point-pinned one-sided-load demonstration from a separately declared, finite-area supported fixture; changing the fixture is not credited as a solver improvement. Neither numerical acceptance nor penalty insensitivity closes the independent spatial-accuracy gate.

An interrupted long output exposed buffered HDF5 metadata loss between saved stations. The serial unified writer now flushes complete frames and atomically maintains an explicitly incomplete index, removed after clean final publication. Abrupt-process-exit and independent-reader checks cover that bounded recovery case; they do not establish SWMR, mid-write/power-loss recovery, or repair of older damaged files. Checkpoints remain the only restart contract.

0.4.3 released composition and next gate (2026-10-17)

Version 0.4.3 was promoted by one composed scientific workflow, not by accumulating isolated feature names. The first bounded native slice is three-dimensional finite-strain J2 with ordinary strong constraints and one analytical rigid plane under frictionless penalty contact. The plane may be fixed or have prescribed translation/rotation. Its contact potential, residual and semismooth tangent enter the same Newton system; the accepted J2 state, contact energy, generalized tool work, reactions and serial checkpoint share one increment lifecycle. The elastic small-strain contact provider must not win dispatch merely because a history material also exposes young and poisson.

This is an experimental serial composition, not a general forming solver. The released bounded route adds one triangulated rigid tool with current-trial BVH projection and normalized prescribed motion, but only while active contact lies on planar facets or coplanar interior seams. Its force--moment--work and joint restart evidence are recorded in the candidate gate. Curved/creased die transitions, periodic MPC reduction, mixed displacement-pressure elements, friction and distributed implicit assembly remain explicit follow-ups. Promote them one at a time only with tangent, force/work/energy, failed-attempt rollback, restart and independent refinement evidence. For MPC, reduce the total residual and tangent including contact; do not bolt an unreduced contact force onto a converged plastic solve.

The tagged Linux wheel and source archive passed the installed-use, scientific, MPI and macOS same-source release ladder and were published as 0.4.3. The candidate record retains the exact evidence and the limits. Future promotions still require reviewed supported/unsupported combinations, independent references, sensitivity checks and installed-platform evidence. A passed unit patch or one successful forming picture is not release certification.

The next experimental slice now has exact active-branch sphere/cylinder curvature, compared against an independently differentiated UFL penalty potential. A separate serial affine-J2 composition reduces the full material-plus-fixed-contact residual and tangent through the same periodic MPC on Newton and line-search trials, with fixed-tool restart and dual evidence. These are source-level candidate results, not a new release claim; see the bounded candidate record. Noncoplanar triangulated transitions, MPI implicit contact, friction and mixed pressure retain separate gates. Measure complete accepted-step cost before choosing an acceleration target.

The follow-on source candidate composes normalized moving rigid contact with serial affine-periodic J2 using the same full-space Operator and accepted increment lifecycle. It adds provider-owned tool force/coordinate work, restartable dual history, pose rollback and checkpoint rejection of incomplete work paths; see the moving affine contact record. This closes the bounded moving-tool/MPC composition gate only. Noncoplanar triangle transitions, MPI implicit assembly, friction, mixed pressure and independent curved-structure validation remain open. The source candidate is not promoted to a released capability by these local tests.

The next serial source candidate removes an unnecessary contact/strong-BC disjointness restriction by lifting the native tangent and retaining constrained matrix rows. A full-versus-quarter indentation comparison passes at equal resolution. Its Hertz half-space screen does not yet pass an independent accuracy gate: finite support, mesh and penalty effects still need separation. See the Dirichlet and Hertz candidate record.

The same source branch now composes multiple named frictionless tools in both the ordinary serial strong-BC J2 Step and the affine-periodic J2 route, with the total residual and tangent reduced by the same MPC. Strong-BC J2 also admits fixed or prescribed-moving analytical-plane Coulomb pairs with separate accepted stick/slip and pose states, a generally nonsymmetric tangent, separate friction loss channels and serial restart. These are bounded source candidates, not a release promotion. Open/nonconvex sharp features, curved or faceted friction, friction with affine MPC, distributed implicit assembly and an external friction benchmark remain separate gates. For rotating planes the current-pose master transport is included in the fixed-old-state tangent; focused FE directional checks, nonzero moment, tool work and restart pass.

A further narrow source candidate admits a closed convex outward-oriented polyhedral rigid tool for serial frictionless finite-J2 contact. Its exact interior support-plane gap gives a deterministic limiting normal and tangent at a noncoplanar feature; it does not smooth a crease or relax the old open triangle rejection. Local Operator and planar-limit global tests pass, but a small structural load path across one noncoplanar feature and its interrupted restart also pass. External sharp-feature comparison, broad mesh/penalty evidence and scalable search remain open. See the closed-convex candidate record.

The bounded serial candidate now also carries a selected-branch Coulomb tangent for that certified closed-convex tool, two independently moving frictional pairs, a two-region J2 material map, atomic restart and raw accepted contact-point export. CG1/CG2 hexahedral contact quadrature can be explicitly refined for discrete convergence checks; defaults stay unchanged because higher rules did not monotonically improve the Hertz force screen. The source-targeted and isolated installed-wheel smoke pass on Apple Silicon, but the closest small-strain Hertz screen still has a 1.4% force discrepancy after a farther-boundary Q2 mesh refinement. Sampling the Hertz pressure on the same trace exposes a much larger pressure-shape difference, and changing the face rule does not improve force and pressure together. Sharp-friction observables remain mesh/penalty sensitive, and external same-input Abaqus results are absent. This work remains an experimental, limited serial contact capability, not local Abaqus parity or a new release. The next promotion needs a frozen external comparator and controlled mesh, domain, penalty and increment limits before broader geometry/material/MPI admission.

A later experimental geometry slice adds a reviewed closed, possibly nonconvex triangular rigid solid for frictionless serial implicit contact. It uses a unique nearest-feature signed distance and an exact local face/edge/vertex derivative; it does not average triangle normals. The separate candidate record states its explicit embedding attestation, O(facets) sign-classification cost, local and ordinary-Step evidence, and unresolved global reentrant-edge, same-input reference, MPI and friction gates. This does not expand the released 0.4.3 contract or certify arbitrary STL/CAD.

An additional reviewed triangular rigid tool may now use the same serial fixed-old-state Coulomb Operator on a current planar face branch. A coplanar two-triangle tool matches the analytical plane under translation and rotation, including tool work and interrupted restart; active free or noncoplanar edges still reject and roll back. This extends useful tool representation without promoting arbitrary sharp-edge friction, distributed contact, or an external Abaqus parity claim.

The next source candidate removes the built-in sphere/cylinder type gate for smooth rigid contact: geometry owns a checked signed-gap differential, while the existing Operator and Procedure remain unchanged. A third-party smooth plane enters the ordinary serial J2 Step and an analytical plane enters the periodic J2 composition. This is an extension seam and local regression evidence, not general CAD or sharp-feature contact validation; see the same moving affine contact record.

The separate performance lane follows skills/agentfem-performance: profile whole accepted increments on a frozen case, change one body/contact/data-flow mechanism at a time, and promote only parity-preserving measured wins. Windows/WSL project UX should make project custody explicit and offer an opt-in native-Linux-filesystem execution location when measured I/O warrants it, without silently moving Windows-side projects. It should surface the effective runtime/native-extension status and test installed-package behavior there. Neither a platform switch nor altered mass/penalty/time controls may be reported as an algorithmic speedup.

The 2026-10-23 Windows onboarding follow-up is a separate, bounded product task: reuse WSL preparation for Ubuntu/mirror and Complete Runtime routes, preserve explicit approval/reboot/account handoffs, verify build identity on resume, and report machine-readable progress. Source and mocked Windows-shell evidence are tracked in the installation guide. The source checkpoint also bounds read-only installer queries, preserves literal Windows paths and interruption evidence, and tests concurrent status publication and failure-record storage faults on PowerShell 5.1/7. The review record distinguishes these passing software contracts from real-machine acceptance. Clean Windows/WorkBuddy installation and actual domestic download availability remain promotion gates; this is not a new solver capability or a speed claim.

The 2026-10-26 follow-up closes the environment-to-first-solve gap in the Ubuntu/mirror route: a source-usable standard-library entry plans first, creates only missing environments through one selected channel, reuses existing environments without implicit package changes, and verifies installed identity, health, protected workspace, a small installed-template solve and result provenance. Each attempt retains separate logs and structured failure evidence. Human-readable Windows status views preserve the existing JSON contract and distinguish approval, restart, private account input, interrupted work and saved acceptance. Local contract tests and documentation checks are complete; the Windows 5.1/7 and Linux installed-entry CI gates are required before integration. These checks do not establish clean-machine WSL/UAC, WorkBuddy or domestic network acceptance, and do not promote experimental contact physics.

Integration on 2026-10-26: PR #123 merged at a862e123 after candidate 67940f6f passed Windows PowerShell 5.1/7 contracts, Linux installed-entry and FEniCSx acceptance, and macOS installed-wheel acceptance. The development branch was removed remotely; local generated website changes were preserved. Clean Windows/WorkBuddy and domestic-network feedback remain separate gates. Development now returns to bounded contact verification: separate normal and in-plane edge resolution before another plastic mesh/path run, then obtain an independent same-input reference. Call-local J2 spectral reuse is already implemented and evidenced in the 2026-10-19 record; do not schedule it again as a missing feature or transfer its speedup to the explicit composite case.

Lean project boundary (2026-10-15)

Checked serial array-mesh import and parent-cell-based exterior regions now remove application-owned NPZ validation and boundary MeshTags construction. Input-node lookup reuses the existing identity-based P1/Q1 helper. Private surface selection and bonding eligibility remain explicit scientific inputs; the reader does not merge nodes or create internal contact surfaces. See array mesh inputs.

Promote reusable execution, trusted native compilation/cache, guarded polar increments and accepted response export into the installed package. Keep private constitutive equations, state mappings and problem geometry in the project. project.execute_step wraps the existing serial transient Step; it does not own a second time loop or choose scientific controls. See project execution. Admission requires installed-wheel parity and accepted-state restart/failure checks. This migration alone makes no new full-run speed or external validation claim. MPI orchestration and general multi-point response export remain outside this bounded interface.

Bounded usability sprint and return to physics (2026-10-15)

The usability scope is runtime-aware non-executing project checks, discoverable result files, explicit observable comparison, concise terminal summaries and an installed real-FEM heat campaign template. See workflow usability. It reuses Project, SimulationResult and Campaign rather than introducing another workflow. Field remapping, a GUI and automatic scientific certification are not part of this sprint.

The next scientific lane is composability of contact, friction and history materials: first audit the existing supported combinations and their accepted state, work/dissipation and restart contracts; then close one demonstrated gap with a small general benchmark. Coulomb state, moving tools and explicit friction already exist below; do not rebuild them. Damage coupling must preserve material energy semantics and failure rollback, not inherit validation from elastic friction tests. Keep bounded thermoelastic coupling as a maintained capability, not a new open-ended rewrite. Shell/solid-shell forming, self-contact and broad industrial validation remain feedback-driven longer-term work.

Installed explicit workflow and delivery gate (2026-10-14)

The bounded serial response-only Hex8 workflow has progressed from its private startup prototype to ordinary model.step, reusable core execution, readable case files and a completed frozen full-grid review path. Native body/auxiliary arithmetic, compiled-material dispatch and Q1 contact acceleration now ship as optional core modules; confidential equations and source geometry stay external. See native explicit execution for the supported interfaces and performance evidence for measured scope.

Frozen candidate 608c50bb has now completed the same full path in 14,646.987 s (4 h 4 min 7 s), 5.29% less than the previous core baseline. Final checkpoint, reporting journal and independent volume/Warp checks pass; this does not close source-case or numerical-inertia validation. The next architecture slices are recorded in portable explicit performance: separate scientific accuracy from execution configuration, target the compiled accepted-increment path, and investigate the controlling time limit before committing to general multirate or device execution.

The delivery priorities are now:

1. Preserve the completed installed candidate and its platform/CI, full-run, response and volume-output evidence. New algorithms are compared at a declared accuracy target, not forced to reproduce its exact step count.
2. Close independent source-case validation and inertia/time/penalty sensitivity without changing inputs to meet a timing target. Completion of the reviewed experiment does not establish equivalence to the source solver's contact, initial-bond or variable-mass behavior.
3. Use whole-increment evidence to choose the next bottleneck. Preserve one accepted State owner; optimize execution/data movement without a second solver or a case-specific core. Keep unsupported distributed response-only combinations rejected until their own lifecycle gates pass.

The follow-up adds a frozen-project preflight with separate integrity, artifact-platform and runtime-readiness results; it must not turn this Mac's binary/worker configuration into Model requirements. Ordered installed comparisons rejected the new copy/scan execution candidate; no new solver optimization was selected in that follow-up. A separate five-times-larger time-step experiment was rejected after early energy growth; material incremental admissibility alone is not global dynamic stability. Next-method work must identify the controlling body/contact/damping bound before selecting a multirate scheme. Do not promote that failed screen or a small copy/scan benchmark as a new full-run speedup.

The older dated entries below document intermediate gates, not the current claim that only a startup pilot exists. They remain useful as provenance for scientific boundaries that have not yet closed.

Corotational history-material gate (2026-10-10)

The source-material review now has a compiled local reference and an independent private batch implementation. General polar-frame kinematics are implemented as a private core candidate, with objective stress transport and independent noncoaxial integration checks. This does not admit that material to finite-Hex model.step: its work-only energy output and response-only stability need explicit contracts, followed by single-element/global lifecycle verification. The next local slice now supplies a genuine response-only work contract, serial quadrature transactions and real Hex8 internal-force assembly without inventing DPDF/SENER fields. Trial failures restore state and response fields. The private response-only route now advances the existing central difference Procedure, records internal work separately from artificial energy, and restores accepted material/force state atomically. Two moving planes or bounded lateral cylinders compose with the existing contact work ledger and disk restart. Independent explicit stability, ordinary Step admission and MPI remain open. The original 255,041-cell mesh now runs a bounded four-tool startup pilot and passes independent continuous/restarted four-increment bitwise comparison. That endpoint is only 4e-10 s, not the full 0.1 s loading. Original interface initialization, variable mass scaling and commercial comparison remain open. Current-configuration deformable Q1 penalty contact now has a private serial face-interior composition with force/moment/virtual-work and time-refinement checks. This does not admit edge/corner normal cones or source ALL EXTERIOR hard contact. A following closed-solid candidate now classifies all 1,000 unbonded source candidates with two true-Q1 rays; they are outside the complete patch, so old open-patch negative gaps were not evidence of penetration. Concave-edge force/moment/work checks and serial two-body restart pass. The original case can opt into this bounded initial-candidate contact; arbitrary ALL EXTERIOR search, collision detection and mass-scaled full loading remain open. Follow-up: private accepted artificial-viscosity work/restart now passes its targeted composition gates; this does not close variable mass scaling or vendor-equivalent viscosity/stability. Explicit reference-surface adjustment proposals preserve the original mesh and require an independently selected bond mask plus validation of affected bulk cells. The source's initial bond set is still unknown. Its body/body hard-contact behavior is also not equivalent to a compression spring. A reusable quintic amplitude and amplitude-driven rigid schedule now address the source loading-history mismatch without changing the existing half-cosine smooth_step API. The ordered nine-gate plan is corotational material development. Keep confidential source equations and parameters in external project providers.

2026-10-11 continuation: segment/BVH pruning reduced the private 1,000-point closed-Q1 query median from 11.812 s to 0.641 s with identical signed gaps. This is a search-component result, not whole-loading throughput. Fixed and variable numerical inertia now have private serial state/identity contracts, accepted acceleration adjustment, artificial energy/linear/angular-momentum inputs, atomic rejection and disk restart. A selective element-bound planner assembles nodal inertia without changing physical constitutive density. The planner consumes supplied positive element bounds; it does not establish an evolving damaged-material/contact/damping stability certificate.

Remaining source-reproduction gates, in execution order:

1. Admit a justified evolving increment/mass policy and quantify added-inertia sensitivity; retain separate negative-curvature/damping handling. The initial elastic screen alone is approximately 8.5e-9 s, not the requested source dt.
2. Extend bounded initial candidates to the required body/body contact coverage with missed-crossing protection; penalty convergence is not exact HARD contact.
3. Confirm source initial bond eligibility/adjustment and review stabilization/ characteristic-length differences. Do not invent interface damage or deletion absent from the supplied source deck.
4. Run nonzero/full physical loading with force--displacement, artificial-energy, penetration, increment and mass-scaling convergence; then compare independent source-solver evidence. Startup/restart capacity is not a completed reproduction.
5. Promote only proven combinations into ordinary Step options and MPI; keep general mortar, relaxation stabilization and deletion as separate product gates rather than prerequisites invented for this one elastic interface.

User-driven mesh/interface slice (2026-10-09)

2026-10-10 continuation: the bounded finite Hex8 route now has ordinary-Step lowering and accepted-response energy sampling, declared virgin response, prescribed-motion work, material dissipation, isotropic reference-interface energy and energy-preserving restart. Independent finite elastic/plastic stretch checks pass, with time-step refinement of the balance residual. A bounded opt-in negative-curvature growth resolution policy preserves signed spectrum evidence; it is not a general nonlinear stability guarantee. Nonuniform plastic response and distributed bulk ownership/failure tests now pass. The shared rank-local numeric checkpoint encoding additionally covers same-partition finite MPI restart, including empty ranks, corrupt or swapped payloads and failed publication. Independent private and ordinary-Step plastic capacity runs now both complete 262,144 cells and 1,000 increments; the public run peaks at 3.20 GiB. A separate three-pair, 4,096-cell plastic comparison reduces median integration time by 38.2% with identical compared response arrays; see performance evidence. This is not an industrial-forming or universal speedup claim. Distributed interface integration, cross-partition finite state recovery, damage/deletion and advanced hourglass laws remain separate gates, not requirements satisfied by these tests.

The tester's Hex8/interface brief is assessed by capability, not by reproducing one private input file:

Requested part	Current evidence-backed boundary	Next missing gate
A1 uniform-gradient elastic Hex8	Ordinary Step; isotropic/rotated orthotropic, stiffness hourglass, lumped mass, bending/wave/patch checks and separate artificial energy	Application-specific accuracy and commercial cross-comparison are not inferred
A2 finite deformation	Bounded history-material Step, finite objectivity/tangents, independent proportional and nonuniform plastic references, owned-cell MPI and same-partition restart	Automatic material-owned stability estimates; viscous/relaxation stabilization; damage/deletion and mass scaling remain unimplemented in this path
B1 nonmatching bonding	Serial fixed coplanar P1 and affine planar Q1 traces, genuine elastic law, coverage/force/moment/work checks and accepted recovery in explicit dynamics	Warped/general surfaces, cross-rank interface ownership and general convected material frames
B2 integration quality	Common-refinement overlap integration retaining both original bases, side-swap and patch evidence	Not dual mortar, arbitrary sliding, re-pairing or post-failure contact
Combined industrial forming	Small finite-bulk/equal-stiffness elastic-bond Step verified, without case-specific solver code	Moving contact plus the actual orthotropic finite material, damage policies and independent industrial reference remain separate work

Prioritize these genuine gaps instead of advertising all of A2 or the complete industrial model as finished. Installed-use and performance evidence support usability and efficiency, but do not promote scientific scope by themselves.

The next bounded stability slice should first expose material-owned evidence at the initial and accepted states, with explicit configuration, units and validity scope; do not infer an isotropic wave model from an arbitrary dP/dF array. Compare the composed bulk/hourglass/interface estimate against the existing fixed-step screen and independent time refinement before replacing the caller ceiling. An independently validated stability response could then permit response-only material updates where the explicit integrator needs no Newton tangent; do not skip tangent generation while the current admission screen still depends on it, and measure the actual whole-step benefit. Automatic variable increments are a subsequent Procedure change: actual accepted times, prescribed-motion work and restart identity must remain consistent, and a rejected attempt must not advance material state. Do not add an alternative explicit solver or silently enable mass scaling to make this gate pass. The tester's actual finite orthotropic material and moving contact combination require their own evidence after these shared contracts.

Continuation candidate: a provider-declared FirstPiolaTangentEnvelope now lowers through the fixed-reference mass-gradient Gram and ordinary Step. The bounded serial route checks its principal-stretch domain, signed tangent and restart identity; independent element/assembled spectra and a Neo-Hookean external provider exercise the contract. This removes hand-entered frequency ceilings only for materials supplying a valid domain envelope. Native finite J2 still needs its caller ceiling; its history-dependent proof and new distributed envelope tests remain open. Variable increments, response-only updates and moving-contact composition are not implied by this candidate.

Contact follow-up: numeric nested material checkpoints and accepted-force sampling now pass serial composition with a moving plane and a finite elastic history provider. The common work ledger, summed spectral ceiling, time refinement, restart equality and post-commit rollback are exercised together. The next candidate lowers one registered frictionless translating plane through ordinary finite-Hex model.step, preserving the separate constitutive owner. It passes a scalar finite-compression ODE comparison and a native finite-J2 plastic work/restart test. Curved/rotating tools, multiple pairs, friction and the tester's actual orthotropic material remain open; do not equate this serial slice with industrial forming completion.

Distributed contact follow-up (without a bonded interface): the same ordinary Step now passes five two-rank gates for serial equivalence, same-partition restart, empty partitions, rank-inconsistent admission/corrupt outer-state rejection and post-commit rollback. The combined contact/finite-Hex MPI regression passes 62 tests per rank; serial contact/bond regressions pass 57 with eight MPI-only skips. These are overlapping selections, not 124 distinct MPI tests. Cross-partition finite-material restart remains separate work.

The next prerequisite hardens the existing sparse cohesive communication owner: collective schedule/payload rejection, idle consumers and transpose virtual-work checks now pass with two/three processes. This preserves sparse numeric payloads; the added readiness exchange carries only small status records.

The following bounded distributed nonmatching slice now owns quadrature points once, uses remote trace schedules, assembles single-count interface energy and the global owner-summed stability bound, and preserves joint same-partition restart. Two-rank elastic/J2 bulk + bond + tool tests compare with serial values, ordinary results and exact resumed trajectories; corrupted interface identity and post-commit failure roll back collectively. Reference geometry is replicated and point identity remains partition-bound: scalable geometry discovery and cross-partition finite-material recovery are NOT completed by this slice.

The next serial combination now composes bulk, an equal-stiffness nonmatching elastic bond and an external translating plane through the same Step. Contact and bond must have disjoint trace nodes (even shared edge nodes are conservatively rejected). All three stiffness contributions enter the existing summed stability bound once. Independent four-mass axial dynamics, temporal refinement, separate energy channels, native-J2 plastic response and exact interrupted recovery are the acceptance tests. This is not post-failure contact or the tester's unspecified finite orthotropic law. Obtain that law, orientation update and parameters before claiming equivalence; do not substitute J2. A separate, rotated orthotropic Green-strain test provider also passes tangent, energy-derivative, objectivity and joint recovery checks without core changes. It proves the extension seam, not equivalence to the private user material.

This slice implements bounded fixed-reference nonmatching cohesive trace integration and uniform-strain Hex8 stabilization. These are general formulation capabilities, not an automatic translation of a specific Abaqus forming model. Reuse triangle search and existing cohesive laws; keep pairing, material state, force assembly and Procedure separate. The bounded design and scientific promotion gates are in nonmatching interfaces. Each geometry and execution route requires its own evidence; local trace tests alone do not enable migration, finite rotations or distributed interface execution. Do not introduce artificial damage thresholds for an elastic-only interface, or describe unstabilized one-point quadrature as Hex8R support.

The fixed coplanar P1 route now has common-refinement integration, per-facet coverage checks, independent nodal-force audits and serial two-block series-compliance evidence. Its experimental elastic model.step provider reuses PETSc solving and Result output. The affine planar Q1 extension retains the original four-node basis and degree-four common-refinement integration; global 1:3 and 2:3 compliance tests exercise independent Hex8 partitions. Global nonmatching damage, distributed interface ownership and general convected anisotropic interface frames remain separate gates; the isotropic finite elastic special case above does not promote them. The experimental uniform-gradient Hex8 policy now has compact batch execution, serial global affine/bending/wave evidence and ordinary explicit model.step() integration. Isotropic/rotated anisotropic elasticity, disjoint material regions, physical/artificial energy separation and serial restart are exercised. Failed explicit increments and auxiliary checkpoint rejection restore the accepted field state. General finite forming beyond the bounded history-material route, damage and distributed interface composition remain separate gates. Natural-load and strong prescribed-motion work reuse the existing dynamic ledger; independent acceleration/oscillator tests exercise work and time refinement. Checkpoints capture accepted endpoint history even between output frames, preventing cumulative-work loss. The elastic Q1 interface now composes with this explicit Step, including an additive spectral stability bound, interface energy, serial restart and failed-commit retry. Warped/nonaffine quadrilateral traces remain rejected. These are bounded small-strain capabilities, not an industrial finite-deformation reproduction. The bulk Hex8 path additionally has two-/four-rank assembly, ordinary-Step work, output and same-partition restart evidence. Nonmatching interface communication and cross-partition restart remain unsupported. Rank-local input failures and work sampling on partitions without constrained nodes now have collective guards.

The Hex8 result path now exports optional constitutive DG0 stress, tensor strain, von Mises stress and physical energy density through the existing transient result lifecycle. Material jumps are retained, artificial stabilization stress is not mixed into material stress, and field provenance records the generating formulation. An installed two-rank candidate exercises this path independently of source-tree imports. Bounded energy-only monitoring and local scatter maps reduce measured run time by about 26% on one 8,192-cell elastic workload; this is a workload-specific measurement, not a general performance guarantee.

Hex8 geometry admission now uses bounded Bernstein subdivision of the triquadratic Jacobian determinant, rejecting hidden inversion and unresolved near-degeneracy before assembly. This does not certify approximation quality or inter-cell non-overlap. The installed candidate also runs the Gmsh independent Q1 bonded-block and Hex8 wave examples; both result manifests pass provenance verification. Scientific evidence remains the separate analytic/convergence checks, not the provenance seal alone.

The 0.4 line is an architectural consolidation, not a feature-count release.

The next Hex8 prerequisite now has a private total-Lagrangian contribution with large-rotation objectivity, energy/tangent and existing J2 batch-protocol checks. The bounded route now lowers through the ordinary model.step() with explicit kinematics="finite_strain", a caller-declared complete-path spectral ceiling, one full-domain history material and declared reference density. A DOLFINx bridge exercises existing quadrature trial/commit/rollback, global force and moment, fixed-old-state tangents and a four-increment J2 patch. An instantaneous nonnegative symmetric-tangent spectral screen is tested separately; unsupported tangents are rejected, not silently symmetrized into a stability claim. The private serial path now reuses the existing explicit Procedure, jointly restores nodal/material state after rejected increments, and passes interrupted versus continuous execution. A nonlinear Hencky bar exceeds 10% extension and shows second-order time refinement against an independent ODE. This does not admit arbitrary J2 trajectories: the signed endpoint screen reports positive and negative curvature separately and requires an explicit bounded-growth policy for negative curvature. It is not a general nonlinear stability guarantee. Accepted work, stored energy, artificial energy and material dissipation now share the existing ledger; optional unsmoothed DG0 views expose the accepted response without reintegration. Next are broader independent references and general objective interface/contact composition. The finite contribution already reuses the elastic nonmatching interface in the equal-three-stiffness special case, with opened-interface objectivity, current moment, joint recovery and post-commit failure checks. This does not admit anisotropic/damaging finite interfaces, contact or distributed interface state. The shared finite-material driver now offers optional validated columnar batches, retaining scalar/ordered providers and the same atomic trial boundary. Native J2 uses its existing constitutive equations through this transport; no new material formulation or provider framework is introduced. Serial numeric-array auxiliary checkpoints avoid JSON expansion and preserve legacy reads. Rank-local version-7 payloads now support same-partition MPI recovery, including empty ranks and collective rejection of corrupt/swapped payloads; cross-partition history migration remains unsupported. Spatial refinement uses independent manufactured Neo-Hookean and proportional finite-J2 problems. The plastic reference includes nonuniform tension/compression and a moving yield front, with a separate body-force derivation rather than the solver residual. Large affine endurance runs are capacity evidence only; they do not replace nonuniform spatial or industrial validation. The remaining promotion gates above are unchanged. Do not infer these gates from local element tests or reuse the small-strain constant spectral bound for a finite-deformation trajectory. Its stable middle layer is:

Model -> Operator -> Procedure -> State -> Result / Verification
                         |
                      Backend

Each owner answers one question:

• Model: what engineering problem is being solved?
• Operator: what mathematical contribution is assembled or applied?
• Procedure: how is the problem advanced and solved?
• State: what accepted and trial quantities must survive?
• Backend: which numerical runtime executes the formulation?
• Result / Verification: what was computed, and why is it usable?

For 0.4.0 and subsequent releases, AgentFEM must keep these boundaries acyclic, preserve one provider-owned lowering route, retain atomic state and result lifecycles, and pass a candidate-bound release ladder. The executable audit is:

python promotion_gate.py --target 0.4-foundation

A complete 0.4 candidate additionally requires complete serial tests, representative two-rank state/nonlinear/output/checkpoint tests, clean installed wheel acceptance, unchanged public examples and compatibility imports, Linux and macOS acceptance, and benchmark evidence for every maturity change. Windows runtime acceptance remains a separate product gate.

0.4.0 release and the next bounded cycle

0.4.0 consolidates the foundation rather than promoting experimental physics. The release pipeline must gate publication on the same immutable wheel's Linux/macOS foundation evidence. See release scope.

After release, prioritize in this order:

1. installation/upgrade, examples, result readability, and actionable user failures; protect projects and results independently of runtime replacement;
2. supported mesh--element--formulation interoperability and early quality diagnostics, with small representative engineering regressions;
3. measured assembly/solver/state/output efficiency under fixed scientific inputs; no speed claim obtained by relaxing tolerances or changing physics;
4. one bounded external-validation track, beginning with the existing finite-strain RVE gaps, without reclassifying incomplete evidence as passed.

Each item needs a reusable owner, a finite acceptance test, and a stop/review condition. New materials, forming, additional backends, or benchmark-specific public APIs are not prerequisites for this consolidation. Keep routine CI targeted; reserve the full installed-wheel ladder for milestones and releases.

Agile maintenance slice before coupling work

Do not turn the installation/mesh/performance priorities into another broad foundation programme. The current bounded slice shares transient operator accounting, fixes evidence loss after prepared-system release/continuation, and reduces allocation in exact boundary-identity checks. It changes neither the public modelling API nor scientific maturity. Targeted serial/restart and two-rank tests are its exit gate; see the reproducible measurement instructions in development testing.

The next bounded implementation is sequential thermal--mechanical coupling on a shared mesh, using existing thermal fields, eigenstrain, EngineeringStep and ordinary providers, not a new coupling god object. Audit the following before implementing a public convenience layer:

1. Field transfer now guards same-mesh/space/component identity and explicitly declared equal time. Copy is exact-space only; interpolation is explicit. Equal array length is not evidence for compatible field transfer.
2. A transferred temperature may change only eigenstrain/RHS, or also material stiffness. Preserve this distinction in the existing time-input contract.
3. Define ownership of accepted source and target states, failed-target recovery, and restart provenance without confusing one-way sequencing with converged two-way coupling.
4. Begin with uniform free thermal expansion and restrained thermal stress; then one spatial-temperature, multi-region regression and two-rank restart.

The same-mesh slice is implemented with nodal field transactions and uniform/ linear-temperature analytical response tests, including two-rank rollback and transfer after thermal restart. Temperature/material dependence retains the ordinary operator lowering; no new global coupling solver is introduced. The multi-region handoff regression now combines two conductivities, regional E(T) and thermal expansion with exact cell-average stress checks. The existing hot-wall example restores accepted heat and recomputes only mechanics; a rejected structural trial rolls back field and transfer evidence together. This is durable upstream reuse, not a combined multiphysics checkpoint. The next two-way slice now has a bounded 3D verification prototype with a reversible heat-feedback Operator, unrelaxed residuals, in-memory joint rollback and an independently written monolithic oracle. See coupling design. The private experiment now verifies inward flux, prescribed-temperature heat reactions, affine prescribed-motion path work, and joint durable recovery. Registered strong boundaries and a bounded ordinary Step provider now reuse the physical Procedure and joint restart, with explicit SI/participant-load contracts and final-snapshot output. Accepted-window progress/checkpoint cadence now reuses the existing reporter and policy, with explicit retention. Field-series output remains explicitly final-snapshot-only. Local installed-wheel serial/MPI acceptance passes for the candidate recorded under evidence/coupling/; public-provider cross-partition recovery now also passes 1 -> 2 -> 1 with prescribed motion and accepted force/work history. PR #90 is merged at 1ecf41e0 after final-head FEniCSx/macOS/docs checks passed. The 0.4.1 release candidate retains the immutable-wheel publication gate. A separate smooth Fourier-mode oracle now shows approximately second-order spatial and first-order temporal convergence of both fields through the ordinary Step; its 1D mode embedded in 3D is not general multidimensional validation. See the coupling design for parameters, errors and fixed acceptance thresholds; no general coupled capability is declared stable.

Execution order: finish this bounded thermal--mechanical product route first, then audit contact against existing moving-tool/work/search implementations before selecting any missing capability. A post-release two-rank audit passes 34 tests per rank for the explicit contact provider, distributed search and accepted work. The serial search microbenchmark confirms existing BVH pruning against exhaustive projection without modifying solver physics. Next measure representative solve-stage costs before optimizing; implicit friction, wider topology and extreme-scale routing remain separate gates. Do not reimplement already tested contact features based on an outdated roadmap. Mesh compatibility and measured performance changes follow demonstrated application needs, not a parallel architecture rewrite. Small releases do not rebuild complete installers.

Next bounded implicit-contact slice (0.4.3 candidate)

The already-verified explicit triangle search is not itself an implicit contact solver. The analytical-plane finite-J2 composition first established one Newton residual/tangent, force/work evidence, rollback, and serial restart. The subsequent fixed-facet candidate reuses RigidContactPair, the reviewed BVH, and the same J2 Procedure, adding a current-trial boundary-quadrature residual and fixed-branch tangent. Its scope and local evidence are recorded in the fixed-facet note. Keep it experimental and serial; do not promote curved/creased tooling, moving-body work, periodic MPC, friction, or MPI by association.

The stacked moving-facet slice adds one normalized rigid pose without rebuilding the reference BVH, plus accepted force--moment work and pose rollback. The follow-on smooth analytical surface and fixed-tool periodic-MPC source candidate is described in the new bounded record. Neither slice certifies noncoplanar triangulated edge transitions or MPI.

Promotion order: (1) close full installed-wheel CI and a candidate-bound scientific comparison for the analytical plane; (2) independently check faceted-contact geometry, force, moment, tangent and convergence under nonuniform deformation; (3) verify prescribed tool motion with generalized work and joint rollback beyond the local plane-limit evidence; (4) reduce both contact residual and tangent through periodic MPC; (5) establish MPI owner/ghost and restart evidence. A new release may ship a clearly bounded experimental subset, but its version number must not imply that the later gates passed. Performance is measured on fixed physics and an immutable installed candidate, separately from changing penalty, inertia, or time discretization.

The current bounded serial candidate has progressed beyond the original single-facet slice: ordinary finite-J2 Steps now exercise prescribed-DOF contact, named multi-tool motion/work, reviewed planar/closed-convex Coulomb branches, restart and accepted-point output. The candidate acceptance map separates these completed source and installed-Mac checks from the remaining external same-input comparison, general sharp-feature laws, distributed implicit contact and cross-platform release gates. Its Hertz force, pressure, mesh, penalty and energy screens are directional diagnostics, not an industrial-contact accuracy certificate. Do not silently widen the public capability merely because this candidate can complete a Newton solve.

The next closed-general candidate extends the serial rigid-tool branch to a caller-reviewed embedded triangle solid. Its normal law retains current-geometry nearest-feature/winding evidence and rejects ambiguous projection; a small ordinary J2 Step now exercises an active reentrant edge, plastic state, tangent check and restart. See the bounded closed-solid record. The Coulomb extension is restricted to active locally planar faces; it rejects physical sharp edges or vertices and rolls back the failed trial. The serial affine-MPC path now has a versioned accepted friction-State checkpoint and a tested plane/triangle/closed-face tangent reduction, but no arbitrary sharp friction law or distributed operator assembly. This is not yet a same-input external comparison, a converged parameter study, frictional sharp-feature law, large-CAD backend or distributed implicit contact. The fixed L-tool screen now exposes a concrete numerical gate: changing only the positive face rule changes which physical sharp-edge points are sampled, and a six-cell penalty bracket remains far from a hard-contact limit. Even the same-input degree-7 screen through twenty cell divisions still changes maximum PEEQ and active edge samples; a nearly stable resultant is not enough. The next frictionless method experiment should partition slave-face integration at zero-gap and nearest-feature transitions while preserving a common residual/tangent trial partition, then compare force, moment, penetration, plastic history and tool work across mesh, penalty and increment refinement. Do not elevate higher polynomial face quadrature or a faster sign classifier as an accuracy repair. Stateful edge friction and distributed ownership follow only after this frictionless integration gate is independently checked. The first bounded slice now clips a current-trial affine P1 slave triangle against an analytical rigid plane, with a positive common residual/tangent rule and an independent narrow-strip oracle. It does not resolve the L-tool's changing nearest feature or replace the still-open mesh/penalty convergence and independent same-input acceptance gates. An analytical two-plane counterexample now fixes the next design constraint: at an active sharp-feature switch the moving internal boundary contributes a nonzero line term to the contact tangent. A zero-gap clipper alone cannot be rebranded as a consistent L-tool edge algorithm. A research-only positive two-plane area-plus-feature-line oracle now matches analytic and directional derivative checks, including oblique planes and a spring-supported global fixture checked against independent integration. A one-tetrahedron linear elastic fixture also equilibrates, with its hand-derived bulk stiffness independently checked against DOLFINx assembly. The remaining nonconvex L-tool gate needs its nonlinear nearest-feature transition law and convergence evidence before any such Step promotion. The bounded convex-support-plane subset is now an opt-in ordinary-Step and serial affine-periodic J2 implementation: it partitions at most four dominant planes per affine slave triangle and includes the compressed feature-line Newton term. A real finite-J2 Step, full Jacobian including MPC reduction, moving-tool work and serial restart pass. Fixed support planes are prepared once; every current trial still recomputes its gap, active subcells and projection, and a projection feature disagreement rejects the trial. An independent symmetric convex-corner fixture now checks exact potential, integrated force and five directional total tangents with three dominant planes. A four-feature rectangular window now additionally passes an independent exact-energy oracle, perturbed directional tangents, assembled FEniCSx tangent and an interrupted finite-J2 restart; five positive-area features still reject. This expands a bounded convex subset, not the nonconvex reentrant-tool convergence claim. This does not close the nonconvex L-tool's nonlinear nearest-feature transition or the independent mesh/penalty benchmark. The current same-source fourteen-cell L-tool penalty screen is an explicit admission blocker: increasing its penalty from 200 through 1000 to 5000 continues to change contact resultant, penetration and fixed-volume plastic response substantially. Agreement between two face quadrature degrees at one penalty is not a substitute for this limit. Before promoting general sharp-tool contact, establish a mesh-by-penalty response matrix with a physical penetration tolerance and an independent same-input reference; retain the bounded convex option as experimental in the meantime. If a penalty-only choice cannot satisfy both penetration and conditioning, evaluate an explicit augmented-Lagrangian normal law as a separate Operator and accepted-multiplier State. Do not silently switch enforcement inside an existing contact pair or reuse the penalty checkpoint identity. The installed-candidate penalty-20000 diagnostic did not complete under either basic or backtracking Newton, even after doubling fixed increments. This is a bounded failure observation, not proof that augmentation will converge or that a higher penalty is universally unusable. First prove the multiplier update and penetration tolerance on a fixed-point planar trace; the trial-partition route has changing quadrature points and needs an explicit State transfer/identity rule before any multiplier history can follow it. The next bounded slice is now implemented: explicit steps.normal_augmentation(...) lowers through ordinary standard J2 with one fixed or prescribed-translating analytical plane, serial P1 tetrahedra and coalesced nodal multipliers. It reuses the existing Newton, material trial state and joint accepted checkpoint. Independent homogeneous limits, plastic loading/unloading, failure rollback and restart have targeted evidence. Redundant strong/contact normal rows reject early. This is still experimental: general multiplier stability, general MPC, rotating/curved tools and partition transfer remain separate gates. The fixed-plane accepted-work gate now records signed numerical contact work, a non-cancelling discrete work bound and opening/closure contributions. The AL functional is never stored energy. Original and contact-aware energy residuals remain visible, with independent cyclic-reference and increment refinement evidence; no automatic energy-verification claim follows. Legacy checkpoints without work history remain unavailable rather than fabricated. The fixed-plane route additionally has a scalar, non-FE coaxial plastic-stop reference through loading, unloading, reverse flow and recontact on two meshes, including restart after reverse plasticity. That check exposed a resolved-amplitude/normalized-coordinate confusion: signed load amplitudes must scale ordinary strong values, and fixed tools must not validate them as motion coordinates. Time-dependent boundary histories and moving-tool ranges keep their own domain checks. This does not enlarge the AL geometry scope. The subsequent translation-only gate retains the same material-side nodal multiplier basis while recomputing gaps against the current tool pose. Tool work is recorded once as external input; numerical relative-gap work is separate. Scalar plastic compression/recontact, increment refinement, frozen-state tangent, joint failure rollback and serial checkpoint checks cover this bounded path. The new public translating-plane example uses the ordinary Step, not a case-owned solver. Rotation, changing normals, MPI and general MPC augmentation remain closed until independently established. The subsequent bounded affine gate admits one serial P1 planar pair only when every reduced normal row has nonzero, disjoint support. It reuses the ordinary affine path with a pre-trial Operator snapshot and a nested outer Procedure; it never commits plastic history during augmentation. A separate hard-obstacle elimination reference on two meshes/two penalties, cyclic open/recontact, successive restart and injected rollback tests establish this discrete slice. Signed normal work is retained at every accepted station independently of cloud-output cadence. Overlapping support, prescribed normal rows, rotating/curved tools, mixed and MPI augmentation still reject; this does not close general multiplier stability or the independent external reference. The optional safeguarded multiplier accelerator is now implemented on that same bounded route. Its reference-area-weighted secant history is owned by the attempted-increment Procedure, never the material or checkpoint State. Default controls are unchanged. Independent pressure, cyclic energy and restart/rollback checks accompany a 32-increment serial example whose repeated median solve time fell 31.8%; this is not evidence for the explicit composite case, curved/MPC augmentation or general commercial-software speed parity. The follow-up acceptance adds a spatially partial plastic contact patch through unloading/recontact, all-station pressure/state/work comparisons, restart and post-commit failure recovery. Both policies also match a separate equality-constrained elastic active-set reference on two meshes. This narrows the acceleration validation gap without admitting new geometry or constraint families. The a62cef23 candidate passed remote Linux tests and macOS installed-wheel foundation acceptance; scientific parity gates remain open. The moving-plane slice now also has a nonuniform Poisson-coupled plastic compression comparison against independently enforced prescribed displacement on three meshes and two penalties. Point-pressure accuracy is checked separately from displacement. Cyclic accepted-point export records Step coordinate and resolved tool amplitude separately; intermediate output and two successive restarts are exercised from an isolated installed wheel. Before any default penalty/globalization tuning, retain full accepted-path field comparisons and the measured failure boundary. Outer augmentation and inner Newton/search work are now separately inspectable; local rollback uses numeric arrays rather than serializing a disk checkpoint every increment. The material-side follow-up shares one call's elastic-trial spectral factors between the J2 return and its consistent tangent, without persistent caches or new State ownership. Fresh paired tests at three mesh sizes reduce the bounded implicit example's solve/result time by 11.5--13.4% while retaining every accepted field and iteration count. This is measured software reuse, separate from numerical-penalty tuning, and is not an explicit-composite full-run claim. Keep the public scalar contract and independent numerical tangent oracle. The fixed-plane route now also has nonuniform manufactured-solution spatial evidence: P1 displacement and gradient errors converge at approximately second and first order, while contact force and nodal pressure errors decrease. This check exposed and fixed a linear-trace coordinate-tabulation roundoff that falsely reported initial overlap on a zero-gap mesh. Original mesh vertices are authoritative; real overlap still rejects at the declared limit. Pre-correction experimental checkpoints with changed trace identities must continue under their original candidate, not bypass the restore guard. The same manufactured implicit-J2 contact screen also found a distinct scaling opportunity in the linear solver, not the contact law: explicit GMRES/GAMG with a recorded rigid-body near-nullspace preserves the tested fields and reduces accepted-Step time on two mesh sizes. Keep direct LU as the robust default while comparing the opt-in policy on independent meshes, materials, constraints and platforms; never silently select it solely from one machine's timing.

Stop after this vertical route has explicit field/response and state evidence. Nonmatching-mesh transfer, broader nonlinear coupling, monolithic Jacobians, and additional physics remain separate subsequent decisions. Installation or mesh defects reported by users still interrupt this ordering when they block an otherwise supported workflow; no speculative installer rewrite is planned.

0.3.8: the 0.4 foundation candidate

0.3.8 is the deliberate consolidation release before 0.4.0. It is promoted by stable ownership and executable contracts, not by adding another material or solver family. Its six product gates are:

1. one mesh--element--function-space compatibility matrix shared by inspection, validation, lowering, result identity, and checkpoint identity;
2. mesh-quality evidence and early, addressable rejection of unsupported topology or formulation combinations;
3. typed time inputs that distinguish right-hand-side changes, operator changes, state changes, and output-only observations before a Procedure chooses reuse or rebuild;
4. provider-owned reaction, force, work, and energy closure for exact MPC, weak constraints, and the first bounded contact route;
5. one MPI result, state-identity, and checkpoint lifecycle across supported procedures, including deterministic cross-rank-count restoration where the capability declares portability;
6. one independently installed extension that adds a material or Procedure through the public provider boundary without modifying AgentFEM core.

Every gate must preserve the stable Model, Operator, Procedure, State, Backend, and Result/Verification ownership inventory. Passing only unit tests or adding new public names is not sufficient evidence for promotion.

Gate 6 now has an executable installed-wheel acceptance path. The reference extension is built as a separate distribution, discovered lazily through the standard entry-point group, and activated only by an ordinary project's declared dependency. extension_gate.py installs the candidate core and extension wheels into an isolated environment, executes the same Model → Procedure → Result/Verification lifecycle used by users, retains the provider identity in the result, verifies the sealed manifest, and proves by before/after hash that the installed AgentFEM package was not patched. The reference material is intentionally only an acceptance fixture; new scientific behavior remains owned by companion or third-party packages. Release-tier CI now runs foundation_gate.py only after the complete serial suite, representative MPI state/nonlinear/output/checkpoint checks, installed-wheel examples, compatibility imports, and external provider gate have passed. The resulting foundation certificate binds those stages to the exact clean commit and wheel digest. The same release workflow then exercises that exact Linux-tested wheel on a hosted macOS runner and performs one final evidence fan-in; promotion can pass only when both immutable platform records match the candidate. This cross-platform job is release-only, so ordinary targeted and core changes do not consume macOS capacity. promotion_gate.py --target 0.4-foundation no longer waits for a record that no workflow can produce or audits one platform before the other exists.

Fresh-agent acceptance remains a separate product gate rather than a seventh foundation abstraction. Release-tier CI now emits an immutable trial bundle bound to the exact wheel, source commit, task and review instructions. A genuinely fresh AI agent must still create, run, verify and explain the model without repair prompts. The acceptance recorder then binds the project source, structured result, transcript and reviewed explanation to that candidate. Deterministic CI may prepare and audit this evidence, but cannot impersonate the unfamiliar agent whose behavior the gate is intended to test.

For 0.3.8, that engineering contract and candidate-bound evidence route are accepted as the release gate; continued zero-intervention runs by unfamiliar users and agents remain field evidence rather than a reason to hold back the foundation release.

Gate 3 now has one typed contract across the supported transient and ordinary incremental nonlinear Procedures. Built-in and custom time inputs retain separate RHS/operator/state/output effects. Linear implicit dynamics and linear implicit-Euler heat transfer use those effects to select safe operator reuse or refresh; nonlinear first-order and load-path Procedures declare per-step or per-attempt residual/tangent assembly and restore the accepted coordinate on rejection; matrix-free Explicit declares per-increment residual evaluation and reports whether a fixed preflight stability bound must remain conservative over an operator-changing path. Transient checkpoint schema v5 and ordinary nonlinear checkpoints bind the complete time-input identity before authorizing restart. The runtime architecture audit now exposes both owned-module counts and deliberately unowned utility roots in addition to cycles and forbidden cross-layer dependencies.

Gate 4 is closed for its declared bounded routes: exact homogeneous rectangular MPC, exact affine-periodic paths, scalar/normal/matrix elastic foundations, and frictionless contact with one analytical rigid plane, fixed or following a prescribed translation/rotation path. Force and work no longer share one overloaded endpoint record. Providers publish ConstraintDualEvidence for force and physical resultants; nonlinear and non-proportional Procedures publish typed ConstraintWorkEvidence only after integrating accepted stations. The homogeneous MPC contract states exact zero constraint work instead of inventing a unit endpoint coordinate, while affine-periodic work retains its full path and integration rule. Engineering verification checks that serialized duals and path work agree with the balance contract. The contact provider owns its potential, residual, tangent, nodal dual distribution, resultant, penetration diagnostics, and contact energy in serial and under two MPI ranks. Its atomic portable restart envelope covers the solution, increment ledger, events, provider-dual history, natural-load work, bulk strain energy, contact potential, and prescribed rigid geometry, with 1-to-2 and 2-to-1 MPI rank-count acceptance. Moving tools publish force--translation and moment--rotation path work. General surface search now has reviewed serial and distributed BVH infrastructure, stable accepted/trial projection State, collective Procedure acceptance, and a geometry-neutral pointwise frictionless penalty law. The local law preserves the same conservative potential and explicitly separates structural residual traction from conjugate prescribed-surface traction. It does not claim a geometry-consistent Newton linearization or assembled path-work evidence. A backend-neutral ContactTrace contract now carries stable point identity, interpolation and explicit reference/current quadrature measure into a checked reference assembly of nodal residual, conservative potential, resultant and tool moment. The assembly accepts only matching projection records and query coordinates. The DOLFINx trace adapter is executable for tagged exterior triangles of tetrahedral meshes and quadrilateral faces of hexahedral meshes with continuous blocked vector CG1 or CG2 displacement. It restricts the cell basis to positive facet quadrature and evaluates the surface Jacobian from complete Lagrange first- or second-order coordinate geometry. Both routes use partition-independent facet/point identity, synchronized ghost values, and collective rejection of unsupported spaces. Standard Q2 hexahedra also pass the positive row-sum mass gate and the complete Explicit contact Procedure. Standard P2 tetrahedra do not: their row-sum lumped mass has non-positive vertex entries, so Explicit fails closed until a reviewed positive mass-lumped high-order simplex element is available. Prism/pyramid mixed-facet adaptation, implicit linearization, deformable master surfaces, free rigid-body dynamics, and undeclared weak formulations remain outside this gate rather than inheriting its evidence.

Reviewed triangle files can enter the same route as scientific geometry assets with canonical identity and source hashes. The explicit contact Backend automatically selects the serial BVH or the partitioned routed MPI BVH; the Model does not own or expose search acceleration. This closes the import-to- search ownership boundary, not CAD healing or general deformable-to-deformable contact.

The first explicit Procedure hand-off is now executable for that narrow DOLFINx slave trace. A composable residual evaluates the current predicted displacement, repeats exact projection, applies the local frictionless penalty law, integrates the trace, and adds only the owner-reduced contact vector to the already assembled bulk residual. Its projection and work trials follow the same increment commit/rollback decision as central difference. MPI-global potential, force/reaction, active-point count and tool moment remain inspectable. One explicit physical-time schedule can drive proportional rigid translation and rotation; its initial and accepted generalized stations produce force--translation plus moment--rotation work and interval power. Checkpoints bind the schedule identity and restore the accepted path while recomputing the memoryless projection. A caller-supplied whole-system stability ceiling remains available. For fixed-normal or piecewise-planar projectors, the same route can instead assemble a conservative contact spectral contribution from the trace, penalties, Coulomb pressure coupling, lumped mass, and MPI ownership, then add it to a supplied unsafed non-contact spectral bound before selecting the time increment. Curved-normal geometric stiffness still requires a reviewed caller ceiling. The same residual now consumes the reviewed serial and routed distributed triangle BVHs. Stable slave-point IDs cross changing rigid facets under finite sliding in serial and at two MPI ranks, while accepted work remains rank-canonical and pure tangential tool motion adds no false normal work. This closes one bounded moving triangulated-tool route, not a shared multi-tool broad phase, arbitrary higher-order topology, deformable contact, or implicit Newton linearization. Multiple bounded pairs already compose through the residual chain, including time propagation, atomic trial decisions, summed energy/work, and nested checkpoint State; the reviewed mixed pair combines one analytical and one triangulated tool without a case-specific solver.

Explicit time-step composition is now a Procedure-owned contract rather than a contact or fracture special case. Body/material, cohesive-interface and contact screens publish named equivalent spectral upper bounds. The Procedure adds all simultaneously active stiffness contributions, converts the sum to one central-difference limit and applies one safety factor. The legacy finite-strain cohesive screen now delegates to this contract and retains its body/interface compatibility fields only as a result view. This closes the known unsafe min(body_dt, interface_dt) composition; formulation-specific element eigenvalue bounds and state-dependent re-estimation remain later accuracy promotions.

Gate 5 now has an executable checkpoint capability contract. Procedures must separately declare the durable payload scope, accepted save boundary, atomic publication, scientific identity, and MPI rank-count portability; a portable scalar sweep ledger can therefore no longer masquerade as a full-field restart. Transient, ordinary and affine nonlinear, harmonic, J2, creep, viscoelastic, user-material, finite-strain J2, and cyclic-fatigue owners expose that contract in their summaries. Cyclic bulk fields have moved from partition-local shards to coordinate-keyed schema v2 and pass both 1-to-2 and 2-to-1 rank-count restoration. The composed global cyclic lifecycle now passes those same migrations while preserving bulk fields, physical-facet-keyed cohesive history, and the cycle-jump ledger before continuing to the uninterrupted reference. The consumer boundary is now implemented: model.step() resolves the exact procedure contract before numerical execution, rejects malformed or unsupported policies with stable AFM-CHECKPOINT-* diagnostics, and rolls back any provider-registered Step objects when lowering fails. The same typed contract is a top-level SimulationResult record, and engineering/release verification checks every emitted checkpoint schema and portability claim against it. Gate 5 is therefore closed for the declared procedures and tested rank-count migrations. Collective high-throughput storage remains a scale optimization, not a missing lifecycle contract.

What is usable today

The installed capability catalog is authoritative. It covers the supported and experimental boundaries for:

• linear solid mechanics, heat transfer, modal and structural dynamics;
• viscoelasticity, J2 plasticity, Chaboche hardening, creep, cohesive response, finite-strain hyperelasticity, and experimental finite-strain J2 workflows;
• multi-material regions, eigenstrain/thermal-strain semantics, material-aware result projection, periodic constraints, and selected Abaqus migration paths;
• mesh import, quality inspection, results, histories, progress, checkpoint/restart, provenance, convergence, and ParaView-oriented output;
• campaigns, scientific datasets, surrogate workflows, external providers, and learned-constitutive contracts.

Inspect the exact installed truth with:

agentfem capabilities
agentfem capabilities --json

Near-term priorities

1. Trusted mechanics

The next scientific promotions focus on depth rather than catalog size:

1. complete external DCB, ENF, and MMB cohesive validation, including unstable propagation control and closed force--work--energy evidence;
2. advance finite-strain J2 through the remaining external RVE comparison. Ordinary strong-boundary J2 now records accepted-path dead-load work, prescribed-motion reaction work, stored energy, cumulative plastic dissipation, and their mechanical residual in the same rollback and checkpoint lifecycle. A four-to-forty increment refinement reduces the reference patch residual from about 9.4% to 9.3e-6, so the ledger exposes path-integration error instead of manufacturing closure. Current- configuration follower pressure now enters the same residual through its UFL-derived external tangent, while its accepted work integrates the configuration-dependent load vectors at both increment ends. Three- dimensional mixed P2/DG0 state now also passes one-to-two and two-to-one MPI checkpoint/continue with exact split primary fields and quadrature state. The external-beam gate is accepted from one executable package identity without relaxing a tolerance: the external curve, mesh, serial/MPI, checkpoint/restart, and 45 -> 90 -> 180 increment contracts all pass. The 90-to-180 RMS and maximum differences are 0.04460% and 0.21498%, within the fixed 0.2% and 0.5% contracts. For the separate Zhang Q2/DPC1 RVE, content-bound candidate and multi-axis audit records now replace caller-declared convergence flags; the scientific runtime, benchmark fixture/driver, executed high-order mesh, region/boundary tags, and periodic equations are bound explicitly. Path and quadrature slices must retain the identical discretization rather than relying on a nominal mesh-size label. A separate Appendix-B oracle evolving \(C_p^{-1}\) matches the provider's stress, plastic strain, inverse plastic metric and elastic energy on a non-coaxial path, so the remaining external gap is no longer attributed to a J2 normalization or state-variable mismatch. A 2026-10-07 source review found a phase-placement discrepancy between prose and Figure 10(a) in the inspected author manuscript. Explicit geometry_source now prevents mixing those models. Figure-based geometry at 2859 cells gives stress/energy/tangent errors of 0.099% / 0.014% / 0.507%, within the unchanged comparison contracts, but the last mesh pair changes energy by 5.3442%; complete promotion remains blocked by energy convergence and remaining figure-specific evidence. The fine candidate has now been reproduced from clean commit 597e8795, with identical stress and energy. A same-mesh 4/6/8-degree quadrature slice passes. Energy-channel attribution assigns 99.8554% of the last mesh change to the pressure constraint defect, not to the condensed channel; this is diagnostic attribution, not a license to switch the acceptance observable. Offline raw-point reconstruction localizes 97.19% of the defect in the worst 5% of cells, almost entirely in the matrix. An explicit interface refinement diagnostic uses 2448 rather than 2859 cells and reduces the defect by 17.30%, retaining the three external comparison checks. This remains a development experiment, not a mesh-convergence certificate or general adaptive-solver capability. A clean same-policy 1755/2448/3543-cell interface-size sequence now keeps all three external comparisons within their contracts. Its final stress and tangent changes pass at 0.06847% / 0.25812%, but primal-energy change remains 0.85369%, above 0.5%; local refinement does not promote global mesh convergence. All raw-point energy reconstructions agree. Result-envelope auditing additionally exposed a self-hashed manifest; the writer now excludes only its own publication path while retaining external artifact verification. Original diagnostic archives retain the old envelope failure explicitly. The 2026-10-08 follow-up separates local from globally scaled refinement: local primal-energy change reaches 0.46575%, but global change remains 0.72236% (above 0.5%). Local stress and global tangent also fail the decreasing-change requirement. Five clean runs verify their output seals; neither complete spatial axis is promoted. See the spatial follow-up. Figure-specific 285-cell lifecycle verification now passes direct serial/MPI response equivalence, both cross-rank restart directions, and a local macro-tangent perturbation audit. A larger 1e-3 perturbation failed and remains recorded; MPI tangent condensation is still unsupported. These coarse lifecycle checks do not promote the fine spatial/formulation gates. See the bounded verification record. The earlier diagnostics below use the prose geometry and must not be interpreted as unresolved constitutive errors or transferred to the figure geometry automatically. Initial 20/40/80-increment and degree-4/6/8 quadrature diagnostics show that path refinement and over-integration do not explain the remaining Table 5 stress, energy and tangent gap. A 315/459/804/1590/2859-cell spatial sequence remains unconverged: at 2859 cells, close to the paper's stated 2823, the published-energy error falls from about 20.5% to 4.21%, while the first-Piola error remains about 5.59%. Element count alone therefore does not reproduce unpublished connectivity or geometry approximation. The fixed-old-state final-gradient differences now agree with the condensed macro tangent to about 1.43e-8, even though that coarse tangent remains about 21.68% from Table 5. This rules out an internally inconsistent Jacobian as the leading explanation without turning a self-consistency check into external validation. A controlled independently remeshed 1x1/1x2/2x1/2x2 curved-geometry family now passes its 1% stress and 5% energy diagnostic contracts (observed maxima 0.7404% and 4.1493%). An independent Q9/DPC1 hand-polynomial and curved-patch oracle passes. Serial/two-rank response equivalence and bidirectional 1-to-2/2-to-1 checkpoint continuation also pass. Exact discrete topology replication and full mesh/formulation convergence remain distinct open questions. The clean-source, content-bound 1e-5/1e-6/1e-7 tangent study now passes: its coarse-to- middle observed order is 1.998, and the two finest finite-difference tangents differ by only 1.36e-8. Blind refinement, parameter tuning, and tolerance relaxation are explicitly excluded;
3. extend the closed bounded MPC/weak/contact evidence contract only through independently verified provider routes; general contact remains a separate scientific promotion rather than a foundation blocker;
4. scale the now executable partition-independent raw integration-point HDF5 result and portable quadrature-checkpoint identity from the compact root-gathered route to independently partitioned extreme-scale output;
5. retain every material and fracture capability at its proven maturity until its independent benchmark and failure tests pass.

The general-contact expansion after 0.3.8 is deliberately ordered by ownership rather than by one forming example. Model assets will describe surfaces, rigid-body motion, contact pairs and laws; Operators will own gap, projection, residual, tangent, force, moment and energy; Procedures will own search/update cadence and acceptance; State will own active projections and history; Result/Verification will own penetration, generalized work and energy closure. The immutable RigidBody asset now binds scientific surface, fixed/prescribed kinematics, reference point, and rank-independent identity without owning search or enforcement. RigidContactPair now binds one slave boundary, rigid body, and scalar frictionless law without taking ownership of search or Procedure stability. Prescribed analytical-plane motion and its accepted force--moment--work contract are now the first executable slice. Next come the solver-neutral projection contract (now executable for analytical planes, circles/spheres and infinite cylinders), reviewed tessellated surfaces (the serial triangle reference projector and fail-closed geometry audit are now executable, together with a deterministic process-local AABB/BVH accelerator and per-query work evidence plus stable barycentric facet coordinates), distributed ownership/search (the stable-ID facet partition and all-gather correctness reference now reproduce the serial oracle under MPI; scalable neighborhood routing now has an oracle-checked two-stage rank-AABB path with aligned integer/float MPI_Alltoallv payloads; owned exterior triangles from first-order DOLFINx tetrahedral meshes now enter the same stable-ID contract, including empty local shards; imported MeshTags and named BoundaryRegion subsets now preserve their physical selection, while higher-order tessellation and measured extreme-scale routing remain), accepted/trial projection State (stable 64-bit contact-point identity, facet-changing finite-slide trials, rollback, and accepted-boundary snapshots are now executable without importing solver or contact-law ownership), then a Procedure-owned search/update cadence (now executable as an every-evaluation lifecycle with collective MPI rejection and explicit increment commit/rollback) and an Operator that consumes the trial projection. Candidate warm starts may optimize broad-phase work later, but exact gaps and normals remain evaluation-local. Accepted discrete-surface projection evidence is now serialized by stable global point ID and restored onto the current MPI partition, including master facet and barycentric coordinates. It remains audit State, not a search cache: projection is deliberately recomputed after restart from portable accepted displacement and schedule time. The explicit residual hand-off now supports fixed or proportional prescribed motion, a mandatory declared stability ceiling, rank-canonical force--moment--work evidence, accepted-path restart, and finite sliding over one serial or routed distributed triangulated tool. Bounded rigid pairs compose with independent identity and summed energy evidence. The first friction foundation is now solver-neutral: a pair can declare penalty Coulomb friction; stable point-keyed tangential State supports objective normal-rotation transport, stick/slip return mapping, trial/commit/rollback and accepted-boundary snapshots; and its response keeps recoverable penalty energy, irreversible sliding dissipation, and separation release distinct. The bounded DOLFINx explicit residual now consumes that contract. It follows the same master material point through prescribed translation/rotation, forms the accepted slave-minus-master increment, assembles normal and tangential residuals, and publishes separate recoverable, dissipated, and separation-release energy channels. Point history is checkpointed by global ID and restored onto the current MPI partition, including empty local shards. The mandatory caller-supplied stability ceiling now has a conservative automatic alternative for fixed-normal and piecewise-planar geometry. It includes normal/tangential penalty stiffness, Coulomb pressure coupling, cross-component contact coupling, real lumped mass, and ghost-to-owner accumulation. It publishes contact and combined spectral evidence, and deliberately rejects the unsafe shortcut of taking the smaller of separately derived body and contact time limits. The external finite-sliding gate is fixed to the public Abaqus/Explicit deformable-body/rigid-surface family rather than a forming-specific solver. Its parameters, two-stage protocol, and force/friction assessor are machine-readable, while the registry explicitly retains protocol_solid_bridge_refinement_accepted_exact_b31_gate_pending. A real deformable tetrahedral-solid bridge now applies the fixed 500-unit resultant, accepts the frictionless preload, atomically transfers it into a new stage, activates mu=0.3, and completes the 0.1-unit rigid slide. Serial and two-rank runs recover 500 normal force, the 150 Coulomb force, exact action--reaction, positive dissipation, valid projections, actual stable-facet crossings over a bounded tessellated tool, and the declared energy tolerance. This is explicitly protocol-level evidence, not a pointwise B31 reproduction. The bridge now also owns separated refinement evidence: one run halves both stage increments at fixed physical duration, then another increases in-plane mesh resolution while retaining the refined increments and duration. The accepted certificate checks dissipation stability, energy non-growth, and facet migration without claiming an observed convergence order from only two levels. Registered rigid pairs enter the ordinary finite-strain Explicit Procedure without user-authored residual plumbing. The Procedure builds reviewed traces, composes body, cohesive, and every contact spectral contribution once, and preserves per-pair force, moment, tool work, potential, friction dissipation, State, and checkpoint evidence. Imported triangle tools and curved CG2 slave geometry now enter this bounded route; Q2 hexahedra complete the Explicit Procedure while standard P2 tetrahedra fail the positive-mass gate. The next independent gate is an element-equivalent external comparison if a reviewed beam-contact formulation becomes available. Distributed multi-tool candidate routing, measured extreme-scale behavior, additional topology, and implicit friction linearization follow that bounded route. STEP remains an optional geometry-adapter input that is repaired and tessellated before the contact core consumes it. Shell/solid-shell forming and self-contact remain later scientific promotions, not hidden extensions of the bounded fixed-plane provider.

The first imported-tool boundary is now executable for triangle-only formats supported by the optional meshio adapter, including STL. Import requires an explicit coordinate scale because STL does not encode a dependable length unit, binds the file SHA-256 as provenance, canonicalizes exact geometry to stable facet identity, and exposes normal reversal only as an explicit user decision. Mixed topology, degenerate/duplicate facets, non-manifold edges and inconsistent winding fail before search. Approximate repair and STEP-to-mesh tessellation remain external preparation capabilities; they do not silently alter the contact Model asset.

2. Mesh and element foundation

Mesh breadth advances by preserved semantics and executable evidence, not by recognizing more connectivity names.

• Maintain verified P1/P2 simplex and tensor-product import, patch, tag, quality, output, and MPI routes.
• Promote prism and pyramid routes only when the active DOLFINx I/O and solve stack preserves the external topology end to end.
• Keep beam, shell, cohesive, reduced-integration, hybrid, and stabilized elements behind dedicated formulations rather than topology aliases.
• Admit mixed-topology solve domains only when regions, assembly, results, checkpointing, and MPI preserve every block.
• Diagnose quality without silently moving nodes or remeshing a scientific input.

3. Composites and manufacturing

Composite development proceeds through reusable mechanics:

1. stable orientation, ply, thickness, orthotropic material, laminate, and failure-quantity semantics;
2. a public fibrous-shell Step with patch, locking, boundary-moment, state, and restart evidence;
3. contact, friction, inter-ply slip, forming controls, and per-layer output;
4. verified mapping of forming orientation, thickness, and defect state into a structural model.

Paper-specific geometries and conclusions remain outside the core.

4. Engineering workflow

• Keep one Step and one result lifecycle across procedures.
• Improve run comparison, artifact opening, imported regions/surfaces/sets, quality diagnostics, and selected Abaqus migration.
• Preserve project data independently of replaceable runtimes.
• Make long solves observable through bounded progress, stability and energy diagnostics, and restartable checkpoints.
• Keep concise human output and complete structured output as separate views of the same scientific record.

5. AI, data, and extensions

• Keep the core free of PyTorch and model-specific neural terminology.
• Let external providers own runtimes, devices, weights, and framework details while AgentFEM owns scientific conventions, state transactions, evidence, and failure semantics.
• Make campaigns resumable, deterministic, auditable, and independent of the physics definition of one case.
• Treat visualizations as views; retain a stable machine-readable scientific result as the data source.
• Keep MCP and other agent entrypoints thin: they operate AgentFEM rather than becoming another solver.

Promotion discipline

Every capability moves through explicit maturity levels:

contract -> experimental -> verified -> validated

• contract: syntax, ownership, and failure behavior are defined;
• experimental: a bounded implementation and regression evidence exist;
• verified: analytical, manufactured, or independent numerical benchmarks demonstrate correctness within a stated applicability domain;
• validated: suitable experimental or field evidence supports the declared physical use.

Performance evidence, a successful example, or a comparison with AgentFEM's own earlier output cannot raise maturity by itself. Unsupported geometry, missing evidence, incompatible constraints, and stale checkpoints fail closed.

Later, after the foundation

The following remain valuable but must not destabilize the current spine:

• broader contact, fracture growth, phase-field, and multiphysics families;
• adaptive refinement and explicit, auditable mesh repair;
• richer CAD and assembly ingestion;
• a second finite-element backend as an architectural pressure test;
• larger graphical workflows built on the same Model/Step/Result contracts.

AgentFEM will not rewrite element tabulation, quadrature, distributed assembly, or linear algebra merely to appear independent of FEniCSx. Its moat is the stable scientific boundary that tells humans and agents what should be computed, how it is advanced, what state is owned, and why the result can be trusted.