Thermal inelasticity and continuation: bounded development plan

This record contains generic software requirements only. No private proposal, personal information, proprietary material data or customer geometry is retained.

Order and gates

1. Maintenance and ordinary use: PR #132 merged; add a checked zero-copy XDMF continuation view. A view is not a restart or a scientific continuity proof. Fix complete-Step restoration if corrupt later payloads can leave earlier fields changed. Keep existing checkpoints and core state ownership.
2. Thermal inelasticity: reuse accepted temperature histories and the physical creep clock; add the monotone Sinh branch to the existing creep integrator. Then compose creep and cyclic plasticity within one local increment, with one accepted state and one complete consistent tangent. Stage structural admission only after local tests, combined state rollback/restart and energy bookkeeping pass. Do not use the rate-independent provider with artificial normalized time to bypass physical-time semantics.
3. Damage/life: start strain-life postprocessing with explicit strain convention, reviewed curve identity and units; keep lifetime assessment distinct from constitutive stiffness degradation. Global K-R damage and mesh regularization remain separate future gates. No crack-length or industrial-life claim.

Simultaneous creep/plasticity design

Retain the existing complete J2/Chaboche return map P and its tangent Cp. At fixed previously accepted histories solve the six-component stress equation

R(sigma) = sigma - P(eps_mech - CE_old - dt flow(sigma), plastic_old) = 0.

The Mandel basis is orthonormal: no hidden engineering-shear scaling. With L = d flow / d sigma, the local Jacobian is I + dt Cp L; the consistent strain tangent solves (I + dt Cp L) C = Cp. Backstress makes the two flow directions generally different, so no common radial direction is assumed. Each line-search probe re-evaluates from old plastic history. Failure returns no new state. This condensed simultaneous solve is not a one-pass split.

Initial boundary: small strain; constant shared elastic moduli; isotropic J2 or Armstrong--Frederick/Chaboche; power-law or Sinh creep; optional shared secant thermal expansion and normalized Arrhenius power-law rate. Temperature-dependent plastic parameters, static backstress recovery, finite strain and contact are not inferred. A material-point prototype does not imply global admission.

Global integration must preserve separate CE/CEEQ and PE/PEEQ/backstress fields, bind every field to checkpoint identity, and include plastic hardening storage, material dissipation and backward-Euler contributions without calling plastic work stored energy. Do not graft the coupled state onto a CE-only checkpoint.

References and independent checks

• MOOSE simultaneous creep/plasticity motivates solving both mechanisms within one increment. Its default finite incremental formulation is not the total-small-strain formulation used here.
• Abaqus creep integration theory distinguishes rate integration from instantaneous plastic response.
• Tests use an independent scalar root for coaxial isotropic hardening, exact n=1 continuous Sinh relaxation under refinement, nonproportional Chaboche histories, fixed-old-state tangent differences, zero-mechanism limits, and forced failures. None is an independent commercial structural comparison.

No new full composite runtime, MPI composition, release or installer claim is part of this record. Local targeted counts are recorded at each verified commit; skipped rank-specific tests are not counted as passed.

Integrated slice

The serial homogeneous 3D composition now reuses the existing implicit creep Step rather than creating a second nonlinear solver. The joined quadrature transaction saves CE/CEEQ and PE/PEEQ/all backstresses. Accepted plastic work, hardening storage, material loss and discrete contributions are separate. Both serial and portable file formats have serial interruption/continuation checks; MPI execution remains rejected. The ordinary examples/thermal_cyclic_creep project completes through agentfem run with computed (not independently validated) result status.

Promotion remains gated on broader structural time/mesh convergence and an independent multiaxial reference. Static backstress recovery and temperature- dependent plastic parameters are not implemented. The mechanical energy ledger excludes imposed thermal transfer; the isolated plastic energy decomposition is not promoted into a full thermal first-law claim.

Local verification: 156 targeted tests passed in 100.98 s across coupled creep/plasticity, Sinh creep, platform/constitutive models, checkpoint capabilities and the generic small-strain Step. The ordinary source CLI example completed 16 accepted increments in approximately 4.4 s; this is a workflow smoke check, not a speed benchmark or an installed-wheel test.

Follow-up checks add an independent continuous uniaxial solution for linear creep plus isotropic plasticity: 16/32/64-step error ratios verify first-order time convergence. A 1/2-subdivision homogeneous patch gives matching reactions and state; it is deliberately not called nonuniform spatial convergence. The broad affected static/dynamic/operator/stateful selection passed 193 tests in 109.69 s. It exposed and now guards against an old duck-typing hazard: having E/nu must not advertise a history material as stateless elasticity. Explicit supplied-operator workflows remain distinct and supported.

The ordinary project now solves homogeneous insulated heat first, transfers accepted temperature frames on the same mesh, and completes the 16-increment mechanical path in approximately 4.2 s. This records a source-CLI workflow check only, not an installed-wheel or thermal feedback/first-law validation. The project declares its required public API, and agentfem check --runtime passes. A separate final selection of flow/platform, continuation output and assessment regressions passed 178 tests in 11.20 s. Counts overlap earlier selections and must not be summed as unique test coverage. Ruff, generated documentation checks, strict MkDocs and the 57-card knowledge check pass. Local REUSE lint was unavailable because that tool is absent in the numerical environment; repository CI retains that gate.

First damage/life assessment slice

The bounded scalar Coffin-Manson-Basquin postprocessor adds explicit calibration range, source, cycle/reversal convention and an uncorrected-mean decision. It reuses rainflow and the existing creep/fatigue interaction; no stiffness damage or life-driven state mutation is introduced. All 68 tests in the strain-life, existing fatigue/fracture and assessment selection passed. The public contract explains scope, theory source and the risk of dwell double-counting. This starts stage three; it does not close global damage, regularization, multiaxial life or material calibration.

Integration review

PR #133 first full Linux run: 5223 passed, 428 skipped, 8 failed. All eight failures were pytest 9.2 rejecting nested approx(list_of_dicts) in older restart tests; they were not reported as scientific passes. The tests now compare complete record sequences exactly, preserving the previous nonnumeric fallback without relaxing tolerances. The six affected files pass 84 tests under local pytest 9.1.1; an isolated pytest 9.2.0 run including strain-life passes 104 tests. The installed scientific environment is unchanged.

Additional assessment guards reject complex histories and distinguish numeric life overflow from no counted damage. CI must rerun on the repair commit; the first run's skipped macOS gate is not acceptance. No automatic merge or release follows from these local results.

Reference-driven follow-up (2026-10-29)

The verification ladder records official test-method catalogue entries, open institutional literature, analytical derivations, numerical errors and the remaining independent-material gate. Only public catalogue scopes were reviewed for ASTM; no full-standard compliance or qualified material fit is claimed.

Three new structural/reference cases check n=1 and n=3 Norton relaxation, the corresponding synthetic time-fraction integral, and a perfect-plastic closed-cycle work/strain-life reference. These use the ordinary Step. The first test attempt found a missing required stress_exponent in the new zero-creep fixture; adding its explicit n=1 setting fixed the fixture without changing the solver. Focused checks passed 41 tests; the expanded creep, coupled plasticity, fatigue and assessment set passed 124 tests in 106.53 s. These counts overlap and are not additive. Ruff, generated documentation, 57 scientific-card checks and strict MkDocs pass.

The review also fixed a decision bug: zero fatigue previously allowed creep damage beyond the supplied diagram because the out-of-range ordinate was zero. Acceptance now requires actual abscissa coverage, with explicit machine-readable scope/domain metadata and invalid imported damage rejection. This changes unsafe assessment decisions, not equilibrium or material history. PR #133's preceding full CI is still running; this local evidence does not replace the latest-commit remote gate.

Resolved histories and nonuniform joint fields (2026-10-29)

PR #133 at 42ed45b6aee2ea79c166ff34794d19ae0b4aa900 completed all applicable remote checks, including FEniCSx and macOS installed-wheel acceptance. The following changes require a new latest-commit CI run.

The public assessment now integrates the reciprocal of a supplied rupture life along explicit, covered dwell intervals. Physical time, stress units, kelvin temperature, original sampling gaps, relation provenance and segment contributions are preserved. Missing coverage, overlapping holds and invalid relation output cannot publish a partial assessment. Temperature histories from the implicit creep Step now declare kelvin. Zero counted fatigue damage serializes to null lifetime with an explicit status, not nonstandard JSON infinity. Rainflow direction detection no longer loses tiny-scale reversals through multiplication underflow; the published MathWorks 14-reversal example provides an external counting reference, not ASTM certification.

Independent nonuniform antiplane fields exercise the ordinary Step. Pure creep gives relative displacement L2 errors 0.2490, 0.06125 and 0.01433 on 2/4/8 cells. With simultaneous active plasticity and nonlinear creep, errors are 0.12862, 0.03200 and 0.007810 on 4/8/16 cells at 128 increments. Exact equilibrium stress and plastic strain are checked independently; 32/64-step peak-creep errors 0.06348/0.03149 match the known backward-Euler sum. These are synthetic proportional isothermal references, not alloy calibration or nonproportional thermal-field validation. The combined reference, material, assessment and existing NAFEMS-cylinder selection passed 135 tests in 205.80 s. After the history-index change, 86 assessment tests passed in 0.24 s. Counts overlap; do not add them. The ordinary thermal/cyclic project also completed 16 accepted increments via source CLI in about 3.9 s, with physical units and computed, not validated, result status.

The first open rupture-data audit retained source/hash/license and rejected premature calibration because temperature/condition metadata are missing and the landing-page/table material labels disagree. Public catalogue scopes and institutional literature guide the next experimental gate; inaccessible full papers are not evidence of having checked their tables. No private proposal or personal details were copied into this repository.

Bounded assessment scaling check

The first history implementation scanned each entire source channel for each dwell. Ordered-time binary searches now select enclosing source intervals and interior sample views, retaining identical complete assessment records. On macOS arm64, Python 3.11.15/NumPy 2.4.6, one process with BLAS/OpenMP threads limited to one, three alternating baseline/indexed measurements after warmup:

Samples/channel; holds	Baseline seconds	Indexed seconds
1,001; 100	0.001855, 0.002078, 0.001902	0.001451, 0.001325, 0.001357
1,000,001; 1,000	3.312518, 3.476882, 3.128787	0.027396, 0.019692, 0.018547

Each hold covers two source intervals. Timing includes the full assessment call but excludes input creation and serialization; every output record is compared exactly outside the timed region. This deliberately exposes sparse holds on long histories, not the cost of densely evaluating every sample. It is not a finite-element or full-simulation speedup. The algorithmic reason to retain the change is removal of the histories-times-holds scan, not a portable fixed speedup promise. No concurrent simulation was running.

Spatial temperature and response validation follow-up

A prescribed Arrhenius temperature profile has an independently integrable joint plastic/creep field. Relative displacement L2 errors are 0.1379916, 0.0346244 and 0.00813975 at 4/8/16 cells; the focused check passed in 76.86 s. It isolates local temperature sampling/rate dependence with zero expansion, not nonproportional thermal cycling or a heat-equation verification.

Review found that the new joint quadrature consumer could cast complex inputs to real and had no full-array validation barrier after successful local calls. Inputs are now checked before dispatch. Returned state, stress, tangent and energy arrays are staged and checked before any live publication. Thirteen focused fault-injection/input checks pass, including a bad third point's tangent, stress, energy or backstress. This is numerical-input atomicity, not a guarantee against arbitrary process or storage failure during assignment.

The ANL-ART-168 Appendix A pages were inspected in the original PDF, including column units and hold-direction definitions. They are a candidate life-test source, not complete pointwise stress/relaxation histories or a qualified AgentFEM material. The NIMS Grade 91 article was then obtained; Tables 7–9 were checked visually against extraction. Their hours/Celsius/MPa and percent-per-hour conventions must not be mistaken for seconds/kelvin and strain-per-second. Published fitted correlations are not independent raw validation data. Its data have not been fitted here. No universal empirical follow-up factor or proprietary design curve was added to core.

Residual-only local backtracking

The condensed joint material's local backtracking previously requested a full plastic algorithmic tangent for every residual-only probe. It now calls the existing linearization="none" branch for those probes; full Newton states and the returned consistent tangent are still computed as before from the same accepted old state. Twenty-four focused local/material and input checks pass. Direct material-point complex strain is explicitly rejected.

An alternating eight-observation prepared 16-increment thermal Chaboche/creep solve_result comparison, after one warmup per method, measured:

Method	Solve/observe seconds
Original local update	9.729661, 9.836265, 9.953227, 9.915063
Residual-only backtracking	9.120214, 9.138382, 9.215313, 9.236852

Median reduction is 7.08%. All physical state arrays and non-timing histories match exactly, and both accept 16 increments. Both methods use the same new quadrature validation barrier; this isolates the local-update change, not a net comparison against an older release. Preparation, compilation and output files are outside the timer. OMP/OpenBLAS/VECLIB requested one thread in this single-process Mac run. This is not a full-composite or portable speed claim; the earlier ordinary-project smoke uses a different fixture and tolerance.

Final affected selection: 149 passed in 267.88 s, covering joint local and global updates, nonuniform mechanical/temperature references, failure atomicity, resolved creep damage, rainflow, strain-life and Sinh creep. Counts overlap the earlier selections. The final source CLI project accepted all 16 increments in about 3.8 s and wrote fields, the final checkpoint and the computed result record. Neither this smoke nor the synthetic reference suite calibrates an alloy or certifies component life.

Independent rotating-flow and knowledge review

Four additional continuous material references pass in 2.82 s: linear isotropic and Chaboche plasticity, each with constant temperature or heating/cooling. The reference prescribes plastic rate and rotating relative stress, integrates the continuous Chaboche backstress convolution analytically, and independently integrates creep strain before reconstructing the imposed total strain. It does not reuse the discrete update. Chaboche creep and plastic axes differ. 32/64/128 increments recover first-order stress, strain and backstress accuracy; the detailed equations and errors are in the verification ladder. This is material-point evidence, not a newly passed structural or life-calibration gate.

Machine-facing review found stale knowledge claims that Sinh remained local only. Those are corrected without extending its tested combinations. The new joint composition has a separate experimental card so power-law MPI/regional claims cannot be inherited accidentally. The resulting 58-card catalog passes import/contract validation, and 18 documentation tests pass. No solver behavior is changed by this knowledge correction.

The existing regional pure-creep Step also passes four series-shear elastic follow-up checks in 9.68 s. Independent continuous relaxation at follow-up ratios 2 and 5 and Norton exponents 1 and 3 matches first-order FE time refinement. The elastic region has exactly zero creep, interface tractions balance, and total-displacement compatibility predicts the creeping strain. This reuses ordinary material regions and coefficient-zero elasticity; no empirical follow-up parameter or new solver is added to production code. It is a shear analogue motivated by the public ANL two-bar mechanism, not an experimental axial-specimen comparison or joint-regional admission.

Final usage review reproduced a new-history guard defect: ten nominal 0.1 s intervals include a floating-point gap 0.10000000000000009 and were rejected by an exact 0.1 s limit. The comparison now has a recorded 1e-12 relative roundoff allowance, retains the measured source gaps, and has no absolute or large-time-origin allowance. Regressions reject real excess, large-origin gaps and tiny-scale undersampling. This does not alter the quadrature or invent additional source samples.

CI budget follow-up

Run 40812073297 on acd46f3b passed the full serial suite: 5315 passed, 428 skipped in 2497.73 s. Installed native kernels (310 tests), the independent provider and several distributed suites also passed. The 45-minute job-wide limit then cancelled the checkpoint-portability step; later MPI/demo gates and dependent macOS acceptance did not complete. This is not a successful whole-CI result or evidence for the later local commits. Logs identify a job budget interruption, not a checkpoint assertion failure. Increase the bounded budget to 60 minutes and print the slowest 20 serial tests without removing any tests or relaxing numerical tolerances. Final HEAD requires a fresh run.