CI: hung Windows pytest leg can still end cancelled with no log (faulthandler_exit_on_timeout not proven) #1032

#1014 added faulthandler_timeout = 900 plus a step-level timeout-minutes: 25 on the "Run tests" / "Run the full suite against the compiled hooks" steps in .github/workflows/tests.yml, so a hung test's traceback and the step-timeout error land in a log that is kept and uploaded as an artifact.

That fix is proven on Linux only: #1022's own round-2 proof (run 38122057767, job 114445674357, a deliberately hanging test with faulthandler_timeout temporarily lowered) shows the faulthandler traceback naming the stuck test in the kept log, followed by ##[error]The action 'Run tests' has timed out after 25 minutes. -- i.e. the step's own timeout-minutes is what keeps the log.

What is NOT proven: the original failure mode #1014 was filed for is a Windows job ending cancelled with no log at all, after the step ran past its own timeout (run 38122057767, job 114418645594: ~33.6 min past a 25-minute timeout, job ended cancelled, no log). #1022 removed a temporary faulthandler_exit_on_timeout = true setting that was meant to make pytest _exit the PROCESS itself well inside the step bound (bypassing whatever held it on Windows), because the round-2 proof above did not show it doing anything: the step timeout fired and ended the step before any process-level exit was observed. Nothing in #1022 demonstrates that the step-level timeout-minutes: 25 reliably ends the step (rather than the whole job going cancelled with no log) on a genuinely hung Windows leg -- only a Linux leg was exercised with a real hang.

What would close this

A CI run (or targeted proof commit, the same pattern #1022 used) on windows-latest with a deliberately hanging test and the step timeout in place, showing either:

• the step ends failed (not the job cancelled) and the log, including the faulthandler traceback, is uploaded as an artifact; or
• it still goes cancelled with no log, in which case faulthandler_exit_on_timeout (or another Windows-specific mechanism) needs to be reinstated and this time shown to work with a genuine hang, not just the 8s proof that only showed the dump firing.

Filed per #1022's own report, which left this gap open rather than claiming it fixed.

[AI-generated]