Affected Repositories

• PyAutoHeart (primary)

Branch Survey

Suggested branch: feature/heart-rehearsal-clock-skew

Reproduction (clean main ea6c1b4)

Using the evidence fixture in tests/test_cloud_validation.py, set the producer's created_at after the integration's created_at and its updated_at before it: collect() returns rehearsal_matched: False (outcome incomplete).

Implementation Steps

1. heart/checks/cloud_validation.py: add _effective_created(run) returning the earlier of created_at and updated_at (a run cannot finish before it was created, so created_at > updated_at means a skewed start stamp); None when either is missing, preserving today's skip.
2. In collect(), use it for the candidate prefilter (created > observed → continue) and for observed = min(observed, created). matching_rehearsal() keeps produced (updated_at) <= integration start as the ordering guard, unchanged.
3. tests/test_cloud_validation.py: skewed start + finished before integrate → matched, pass, report ts not later than the integrate start; skewed start + finished after integrate → incomplete; existing tests unchanged and green.
4. Full Heart pytest; post-merge, the next Heart Health cloud tick re-judges the live chain. No evidence, threshold or scoring edits.

Key Files

• heart/checks/cloud_validation.py — rehearsal/integrate pairing for the cloud board
• tests/test_cloud_validation.py — provenance and freshness tests

Cloud validation discards a matching rehearsal when GitHub reports a skewed run created_at

Type: bug
Target: PyAutoHeart
Repos:

• PyAutoHeart
  Themes:
• release-validation
• heart
  Difficulty: easy
  Autonomy: supervised
  Priority: high
  Status: draft
  Consequence: glance
  Witness: a heart/checks/cloud_validation.collect() test with injected fetch/get_artifact reproduces run 41290559836's timestamps (created_at 09:50:55Z, updated_at 09:09:27Z) against an integrate starting 09:09:36Z: it matches the rehearsal (rehearsal_matched: true) and the report is not stamped later than the integrate start; a rehearsal whose jobs genuinely completed after the integrate started is still rejected.
  Review-minutes: 10

Source: /health dashboard review, 2026-11-01 (claude.ai/code session).

Why

On 2026-11-01 (US DST-end day) the published board went STALE with
"release validation incomplete: no rehearsal for current source", and
validation_report:coverage read "incomplete — no rehearsal evidence —
v2026.11.1.1.dev85301", although the nightly chain had produced a valid pair:

• PyAutoHands release.yml run 41290559836 (rehearsal, run 853 →
  v2026.11.1.1.dev85301): API reports created_at / run_started_at
  2026-11-01T09:50:55Z but updated_at 09:09:27Z. Its jobs report
  started_at 09:51:10Z and completed_at 09:00:08–09:09:20Z — every
  start stamp is one hour late; the run really finished at 09:09:27Z.
• PyAutoHeart Release Integrate 41291876649 started 09:09:36Z and passed.

collect() prefilters candidates with created > observed → continue, so
the skewed created_at (09:50:55 > 09:09:36) discarded the rehearsal before
matching_rehearsal() (version, run_id, run_attempt, build_sha,
produced <= start) was ever consulted. The release verdict stayed STALE for a
genuinely validated source. Re-running the chain clears it only by luck of
timestamps; the gate should not depend on GitHub's start stamps being sane.

Re-read the API timestamps of runs 41290559836 and 41291876649 before acting —
GitHub may have corrected them.

What

• Order candidates by when the producer finished, not by created_at: the
  existing produced = updated_at <= integration start check in
  matching_rehearsal() is the real ordering guard. Treat a run whose
  created_at is later than its updated_at as clock-skewed (use the earlier
  of the two, or drop the created > observed prefilter and rely on
  produced <= start).
• Do not let observed = min(observed, created) stamp the report from a skewed
  created_at.
• Keep every identity check (version, mode, index, run_id, run_attempt,
  build_sha, completed status) unchanged; never accept a rehearsal that really
  completed after the integrate started.
• Regression tests via collect()'s injectable fetch / get_artifact.

Never clear the verdict by editing evidence or thresholds; the fix is to the
collector's time ordering only.