[stefan-jansen]

Closes #99.

Run one complete installed-wheel suite per supported OS/Python cell, retaining the existing build-once artifact, full matrix, version-specific type checks, coverage thresholds, security, executable/strict docs and final package gate. Python3.15 remains independently owned by PR93.

Separate targeted stress runs on affected native/core/test/qualification changes, weekly and by explicit dispatch. It builds one immutable stress candidate, derives its matrix from the normal CI configuration, verifies commit/tree/hashes and installs that wheel in every cell. Each cell runs ten fresh MACD processes with alternating test order, public feature-order controls and twelve cold-call regressions with fresh interpreters/empty Numba caches. Cold-call tests now compare numerical arrays and NaN positions, not only distinct-value counts.

The installed-candidate check verifies distribution version, actual import origin and every package file against the wheel. Permanent tests install real wheels in an isolated environment and reject actual checkout shadowing, a different wheel version and modified installed bytes. Every installed check also starts a fresh interpreter with this checkout on PYTHONPATH and requires rejection.

Executed evidence

• On base f5cd062, the actual YAML test shell starts ten real pytest sessions. The new one-session acceptance fails, with eight actual failure controls passing: 1failed/8passed in9.77s. The replacement invocation passes.
• Both the configured routine command and native driver reject real pytest failure, collection/usage/internal errors, empty collection, os.abort, a real timeout and passing-output-then-abort. Temporarily changing the producing native runner to ignore process status makes all eight negative controls fail while its positive control passes. Restoring check=True returns18passed. No mocked process status or stdout-success inference.
• 51 focused command/release/artifact cases passed in 17.43s at the final test tree (committed as 3a26a94). Fresh full source suite at clean 3a26a94: 4401 passed/57 deselected in 109.83s. Hosted gates remain required.
• Actual targeted run at f30467d: ten fresh MACD processes each53passed; both public feature-order processes each3passed; all12cold-call numerical cases passed in41.68s.
• Actual minimal17-package environment installs the hosted ff94043 candidate. The previous bare import accepts real checkout contamination; the new verifier rejects it and positively verifies all199installed package files.
• Full Ruff/format/ty, actionlint, locked resolution, all-file hooks, package build/import, Twine metadata and strict MkDocs pass. Full native/platform hosted checks remain required at this PR head before merge.
• Hosted Windows 3.13 and 3.14 source runs exposed an actual harness defect: bare bash invoked WSL rather than the Git Bash used by Actions. At fde92ea, discover Bash from the actual Git installation on Windows and use forward-slash quoted command paths. All 51 focused cases pass locally in 13.87s, with Ruff/format/ty/actionlint/hooks passing. Fresh hosted Windows positives and every failure marker remain required; the old failures are not waived or counted as expected success.

This does not change a numerical feature definition, runtime dependency or release version, waive an existing candidate's checks, alter protection, publish a package or deploy documentation. Hurst prerequisite PR97 is already merged and included normally; this diff is limited to CI qualification and its behavioral controls.

AI assistance: Codex inspected the original policy/crash evidence, implemented and reviewed these changes, executed the listed real regressions and source/artifact checks, and prepared this submission. No independent human or quota-exhausted bot review is claimed.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.