History-material integration acceptance — 2026-10-31

CI budget correction

Run 41218586594 at 5a038155 completed 5909 serial tests with 427 skips in 49m09s. Installed ordinary projects, native kernels, external provider, distributed checks, cross-rank checkpoint recovery and release demos then passed. At exactly the job's 60-minute bound, preparation of the final fresh-agent trial artifact was cancelled. macOS acceptance consequently did not run. This is not full CI acceptance and does not authorize merge.

The release job budget is now 90 minutes, retaining every test, timeout and acceptance gate inside the job. The latest revision must complete CI before merge; previous successful substeps are evidence, not substitutes for that gate. Separating long scientific references into independent jobs remains a future CI scheduling improvement, not a reason to omit them here.

Bounded engineering closure

Keep the five ownership boundaries and the existing public model.step route. No new material family, solver loop, output framework or broader MPI/contact admission is introduced. Integrate the already reviewed thermal/history/plane- stress/checkpoint sequence, then the output-control slice, only on accepted CI. The published 0.4.4 tag stays unchanged; moving source now identifies itself as 0.4.5.dev0 with a development contract. New workflows remain experimental.

The audit found three concrete reliability gaps:

1. solve(until=...) could silently convert booleans, complex NumPy scalars or one-element arrays. Both solve and result publication now require a finite real scalar, excluding boolean coordinates, before mutating state or output.
2. Translating the reference mesh preserves an affine gradient but invalidates construction/checkpoint identity. Hash reference coordinates and geometry dofmaps at construction and check at existing accepted-state boundaries. This is a fixed-reference total-Lagrangian contract, not ALE or remeshing.
3. The version check was skipped by shallow CI checkouts, allowing new source to identify as an already published release. The numerical gate now fetches tag history, and the development contract explicitly retains earlier limits.

Installed project, not source-only acceptance

tools/check_installed_history_project.py is a reproducible engineering gate, not part of the user case. It rejects source/editable imports, copies just the four example files, removes inherited PYTHONPATH, and invokes the installed CLI with the same interpreter. The Linux numerical CI runs it after installing the candidate and the optional independent VTK reader.

For both educational laws, the gate runs planned pause, restart, continuous and sparse-output paths. It compares every checkpoint component and common field, checks the two-file scheduled retention policy, joins accepted segments and warps all 16 times through VTK. Sparse output retains four frames without changing the final state. These exact comparisons concern the same host and runtime, not bitwise portability across hardware.

The isolated installed candidate passed all eight paths locally. Its package, source and checker hashes are retained in the confirmation record. The disposable environment shares the existing FEniCSx dependencies; this does not claim a clean dependency installation, Windows acceptance, MPI support or calibration of the synthetic materials. No full installer was built or released.

Test boundary audit

The first focused run passed 219 material/checkpoint/output tests. A subsequent expanded run passed 540 tests and skipped three explicitly two-rank tests before an interruption-output test failed. Its child interpreter loaded an unrelated installed 0.4.2.dev0, whereas the parent pytest used the checkout. That installed writer lacked the current per-frame flush. The test now selects and asserts the same package path as its parent, for both source and installed-wheel CI. No HDF5 integrity or partial-index assertions were relaxed. The failed run is not counted as a passing expanded suite; final rerun/CI must close that gate.

The 2026-11-01 continuation found the same unbound child-package risk in the live-output interruption test. That child now also asserts the parent's exact package path; a startup timeout includes captured output rather than discarding the diagnostic. The first retry passed 559 tests but timed out at that child; MPI finalization also reported an OFI error on VPN interface utun4. It is not counted as a pass, and the timeout alone does not establish a writer defect. With package binding fixed and process-local FI_PROVIDER=tcp (no global MPI configuration change), the output-only regression passed 67 tests in 3.19 s. The complete affected regression plus documentation tests then passed 604 tests, with the same three explicit two-rank skips, in 77.45 s. Numerical thread limits were one; deprecation warnings were errors. No assertion or scientific tolerance was relaxed. This closes the local regression gate, not the final PR CI gate or distributed history-material admission.

PRs #133 and #134 were already merged. After fresh latest-head CI and dependency checks, #135 and #136 were merged on 2026-10-31 UTC (2026-11-01 locally); the main tree at 7e6e8216 exactly matches the reviewed #136 source tree. The output slice still requires its own CI before merging. The working checkout was moved from temporary storage into the persistent development workspace after restoring missing tracked files from its committed index; existing file contents were not overwritten.

Packaging and CI-scope checks passed 77 tests; repository correctness lint, generated documentation and 58 scientific function cards validate. Local sandbox MPI initialization failures and intermittent execution-approval timeouts are environment failures, not passing tests or evidence of a solver defect.

The previous short-path output-preparation measurements retain their original source identity. This integration adds state guards and makes no new runtime percentage or full-composite acceleration claim.