[renmengye]

Follow-up to #573 (stacked on it; retargets to main when #573 merges). Two defects found after #573 reached its review cap:

1. Wall-clock timestamps. The live test on Per-run trajectory: a kernel-written, hash-chained event timeline #573 showed pr.open with a ts earlier than session.start: attempt.py and the tick passed at=now, the caller's tick time, so events recorded late in a long attempt carried its start time. These calls now record the time of the event. Test: the ending event's ts is wall-clock time.
2. Ending ref matches the kept report. finish_run hashes report.md into the ending event, but the steward wrote its no-change and error reports after calling finish_run, so the ref described a placeholder that was then replaced (round 4 on Per-run trajectory: a kernel-written, hash-chained event timeline #573). The steward now writes its report first; finish_run keeps an existing report, so the final file is unchanged and the hash matches it.

Built and reviewed by Claude; codex review pending.

[github-actions]

Round 1 — reviewed head d9efd5ac — reviewer hermes/gpt-5.6-terra.

terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.

Verdict: no defects found.

[github-actions]

Round 2 (re-run on the same head) — reviewed head d9efd5ac — reviewer summarizer:hermes/gpt-5.6-terra over coverage+credentials+deployment+general+lifecycle+prose.

terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.

Verdict: nothing blocking — 2 advisory notes.

1 finding attached to the lines below.

Advisory (non-blocking):

• Suggestion: The steward report hash fix has no regression test. [coverage+prose] The steward terminal tests only check outcomes and record state, and never read the ending event report digest, so moving either report write back after finish_run would still pass while the ending event again hashes a placeholder report. (tests/test_steward.py:588; high confidence)

Two advisory test gaps: timestamp coverage excludes measure, syscall.decision, pr.open, and wake; and the steward ending-event report digest is not asserted. Rejected as duplicate: the coverage finding at src/outerloop/steward.py:524 and the prose finding at tests/test_steward.py:588 describe the same missing steward report-hash regression test; they were merged, using the latter location and the combined evidence.