Diagnose helper exit-evidence and repo-guard timeouts during full-unit validation #4512

Current execution status — 2026-11-13

Open validation diagnosis for #4426. The original helper assertion reproduced on unchanged base; the candidate comparison instead failed a different guard timeout. Cause and base repeatability remain unproven. Failed receipts remain failed; no timeout or test-policy weakening.

Latest detailed Current state.

Parent: #4368. This diagnosis blocks validation for #4426. It does not authorize unrelated PR maintenance or weakening test policy.

Observed behavior

The exact affected gate at candidate c7bb1990e541174c147e25b9f907d2d9c1625ada failed its full-unit fallback. Related tests passed 376. Unit tests reported 19,756 passed, one failed and three skipped. The checkout stayed clean; the actual runner receipt reports exactCommit: true and checkoutChanged: false.

Failure: packages/platform-android/src/__tests__/snapshot-helper-exit-result.test.ts:873:10, a helper that prints chunks and then times out in the executor leaks no stream. The exit-evidence assertion expected statusCode === 1 and received undefined.

A controlled comparison then ran this identical command once per checkout, strictly sequentially, with no competing audit checks:

AGENT_DEVICE_VITEST_MAX_WORKERS=4 npx --yes [email protected] run check:unit

• Unchanged base 80075b6412dfa94fd765a36407270e21adce8981: exit 1, 19,749 passed, two failed, three skipped. The same helper test failed at the same location with undefined !== 1. A second failure was test/repo-guards/macos-helper-package-contents.test.ts:65:1, timing out after 5,000 ms.
• Candidate: exit 1, 19,756 passed, one failed, three skipped. The helper assertion passed in this run. Instead, test/repo-guards/snapshot-producer-emitters.test.ts:186:1, the scan fails an emitter that names an unlisted producer, timed out after 5,000 ms.

Neither run reached the later temporary-directory-leak or smoke sub-stages. Both checkouts stayed clean. Pinned setup passed. The candidate changes only the Android emulator-lifecycle module and its tests; the original failing helper cluster is unchanged across the comparison range.

Interpretation

The original assertion can fail without the candidate patch. The candidate still has failed validation. The changing failure set does not establish host contention, an infrastructure fault, or the cause of any failure. There is only one unchanged-base observation; it does not establish that base's repeatability. Historical receipts remain failed.

Required diagnosis and completion

Use bounded phase evidence at the existing helper/executor and repository-scan owners. Distinguish child startup/output/termination, exit-evidence normalization, scan work and teardown. If a concrete invariant defect is found, fix its owner and demonstrate detection with unchanged-base or isolated mutation evidence. If the failure does not recur, retain the failed logs and identify the evidence needed at the next occurrence.

Keep meaningful exit-evidence, stream-cleanup and producer-completeness assertions. Preserve #4050's full-unit fallback, worker and isolation policy. Do not increase timeouts, add retries or allowlists, delete tests, change the pool, or run repeated broad gates to obtain a pass. A passing focused case does not clear the failed affected gate. Native Android boot proof remains a separate #4426 requirement.

Evidence is retained under /tmp/agent-device-audit-review/coordinator/next4-apex/4426-current-exact/ (actual receipt receipt-stdout.json) and 4426-unit-base-comparison/ (report.md, base.log, head.log, setup/state records). The failure excerpts above provide remotely readable evidence; publish fuller excerpts with the eventual PR.