ci: separate targeted stress from full installed-wheel matrix qualification #99

Problem

Routine PR/main CI repeats the complete installed-wheel suite ten times in every supported OS/Python cell. The nine-cell matrix therefore runs 90 complete wheel suites per revision, in addition to source qualification and coverage. Python 3.15 expansion would raise this to 120. The repetition count is an inherited engineering choice introduced in 96d62ce, not a numeric acceptance requirement in the shared release/compatibility standard.

Actual evidence

PR92 CI https://github.com/ml4t/engineer/actions/runs/39040904149 qualified the exact candidate and passed every cell. Linux312 job117430141587 reports ten complete 4,378-passed/57-deselected suites: first205.64s, subsequent115.58-119.75s. The repeated whole-suite work increases CI duration and runner demand without replacing focused reproduction of a crash or order dependency.

The original repetition followed a genuine Numba MACD out-of-bounds write and allocator corruption. Preserve its regression and targeted fresh-process stress; do not infer a successful process exit from a passing stdout summary.

Accepted scope

• Run one complete installed-wheel suite in every supported OS/Python cell. Keep the full supported matrix and the existing prerelease policy; add final315 only through its qualified owning work.
• Preserve build-once immutable wheel/sdist identity, commit/tree/version/digests, fresh installation, actual import origin, locked dependencies, normal-JIT behavior, version-specific types, coverage thresholds, security/dependency review, strict/executable docs and final artifact gates.
• Run targeted repeated fresh-process stress separately for the actual Numba/MACD/order/cold-call failure modes, including affected native changes and scheduled or explicit stress runs.
• Prove the configured invocation propagates a real failing test, collection/usage/internal error, empty collection, crash/abnormal exit and timeout. A passing-summary-then-crash case must fail. Preserve positive installed-package and wrong-artifact/revision controls. Workflow-text assertions alone are not behavioral proof.
• Keep protected check names and final dependencies. Do not mask failures, skip supported cells, bypass protection or retroactively waive an existing candidate's configured checks.

AI assistance

Codex inspected the actual workflow history, accepted standards, retained native-crash evidence and current hosted job logs, and prepared this focused maintenance requirement. The proposed source change and new behavioral controls are not yet implemented or qualified.