repository-validation-real-process-2675 signal test fails deterministically on macos-latest runners (blocks all PR unit matrices)

[zaxbysauce]

Summary

tests/unit/scripts/ci/repository-validation-real-process-2675.test.ts — test "records a POSIX signal as crashed, distinct from timeout" (line ~104-122) fails deterministically on macos-latest CI runners as of 2026-09-12. The spawned CRASH_FIXTURE child survives the delivered POSIX signal and exits 0, so the harness records status: "passed" instead of "crashed":

Expected: "crashed"
Received: "passed"
at tests/unit/scripts/ci/repository-validation-real-process-2675.test.ts:117:27

Evidence (9/9 failures across two unrelated PRs, different shards + runner machines)

• PR fix(repo-map): connect the six remaining audit actions + VALID_ACTIONS consumer ratchet (#2540) #2735 (branch fix/2540-repo-map-actions): run 34715696860 attempt 1 unit (macos-latest, 5) 3/3 attempts failed; --failed rerun attempt 2 unit (macos-latest, 5) 3/3 attempts failed. Neither the test, the harness (scripts/ci/repository-validation*), nor any signal-handling code is in that diff.
• PR fix(test-runner,mutation): bounded multi-source scope + analyzer-derived mutation selection (#2492) #2736 (branch fix/2492-test-mutation-scope): run 34716327458 attempt 1 unit (macos-latest, 2) 3/3 attempts failed; --failed rerun attempt 2 unit (macos-latest, 2) 3/3 attempts failed. Same — file absent from that diff.
• Ubuntu shards pass on both PRs (all ubuntu unit shards green). Windows is already excluded via test.skipIf(process.platform === 'win32').
• The flake was first observed on PR fix(summaries): durable no-overwrite summary identity across restarts #2722 (run 34670316494, macos-latest-4) and again on PR fix(memory): serialize cohort destination migration under its live lock #2734 today (macos shard, different file passing there), but as of 2026-09-12 evening it is deterministic under PR-run load (9/9), not intermittent.

Likely root cause

A macOS runner-image and/or Bun 1.3.13 behavior change: the crash fixture (child process expected to die from a delivered POSIX signal) now survives the signal on macos-latest and exits 0. The test already win32-skips; macOS needs the same treatment until the fixture is made signal-robust (e.g. deliver SIGKILL instead of a trappable signal, or detect survival and fail with an explicit "child survived signal" diagnostic).

Interim quarantine

Both affected PRs will add the test file to scripts/ci/quarantined-tests-macos.txt (OWNER: this issue, EXPIRY: 2026-09-26) so the unit matrix is unblocked. Ubuntu coverage is unaffected. Retirement criterion: fix the fixture/harness signal handling, verify the test passes on macos-latest, remove the quarantine entry.