ReviewOS

also looking at this

stacks/drone

test: retain execution lifetime failure observations

#62
Draft chris wants to merge codex/execution-lifetime-observations into main
10 files +174 -64
8 days ago

Native execution-lifetime failures previously discarded their temporary fixture and did not identify which operation consumed the budget. Retain failed owned fixtures, scalar child/runtime/phase observations and settled process status. Evidence-file failures cannot replace the original child verdict. CI archives only explicitly allowed private diagnostic files, excluding linked dependencies.

Current head: febdddd196be66ba6c0ecc2ddbbe20093c8d10a6. Parent and child report their own runtime identities. The test-only setup observer distinguishes native database, account, RBAC and token work from actual producer and retention waits. Production authority policy and the original 2-second native, 60-second retention, 90-second child and 100-second test limits remain unchanged.

Validation and limits:

  • Pickier and application/runtime types pass. All three native lifetime modes pass on pinned Bun 1.4.2 using the new setup observations, as do five shared handover/execution regressions. Separate Spec and Standards reviews report no actionable defects.
  • Real external SIGKILL plus unavailable artifact paths preserves the child verdict and diagnostics. A separate real SIGSTOP campaign exercises the unchanged 90-second watchdog; both deliberately fail their subject test and pass only their diagnostic predicates.
  • The earlier pinned normal run at 16458e5 remains a retained 2-pass/1-fail result. Its completed case spent about 49 seconds in setup/producer work before entering the actual 60-second retention wait, then exhausted the 90-second watchdog. A later instrumented completed case passed with setup 1.42 seconds, database 0.75 seconds and producer 1.26 seconds. These observations do not explain the slowdown or establish the original issue #60 cause.
  • Prior-head CI 39810221070 is terminal FAILURE: the full check, firmware and namespace jobs pass; both browser jobs fail. Their original logs and archives are retained under issues #44 and #63. No canceled run or passing retry substitutes for that result.

This remains a draft diagnostic change. Current-head CI 39819353675 is terminal FAILURE: the check job stopped at the existing 20-second cold-build setup hook before the full suite ran (issue #65). Firmware, namespace and both browser jobs passed. This does not qualify the lifetime suite. The original causal investigations remain open. No physical runtime, individually joined Worker thread or completed hardware qualification is claimed.

10 changed files on the files tab, with 0 review threads.

Sign in to comment on this pull request.