fix(ci-status): wait for the re-run of a failed full run before carrying its failure

[kyle-sexton]

Closes #646

Summary

In carry-forward mode, ci-status ended its wait as soon as the newest ci-lanes status was success, failure or error. After a full run failed and was re-run, a contract-only event on the same SHA read the old failure at once and went red. It did not wait for the re-run. Reproduced in melodic-software/claude-code-plugins#4670: contract-only run 36675804726 went red at 06:00:01 on the 05:56:46 failure of run 36673988526, while that run's attempt 2 (started 05:57:10) was in flight. Attempt 2 passed at 06:12:49.

The failure message also only ever said re-run the full workflow. Once ci-lanes is success, re-running the red contract-only run is enough.

Fix

• A settled failure or error is held open only while its writer re-runs. read_carried_state now also reads the chosen entry's created_at and the run id its target_url names. The status stays open, and the wait narrows to that one run, when that run is in flight with a run_started_at later than the status's created_at. A success still ends the wait at once.
• No mutual wait between contract-only siblings. Only a full run writes the status, so a contract-only sibling is never the writer. Two contract-only runs on a failed SHA still stop at once. A run_attempt > 1 test was rejected because it also matches a re-run contract-only run. A jobs-API discriminator was already rejected in fix(ci-status): wait on any in-flight sibling before reading ci-lanes #562.
• Start-time term. It keeps the writer's own first attempt out of the wait: that attempt writes the status a moment before it completes.
• Known limit. A re-run of a different full run on the same SHA does not hold the failure open. That case fails at once, as before.
• Fail-closed paths are unchanged. A writer re-run slower than the ceiling runs to it and fails. A timeout, a missing status or a skipped lane never passes.
• Both remedies in one message. Every carry-forward red goes through fail_carry_forward. It keeps re-run the full workflow and appends Once ci-lanes on <sha> is success, re-run this run instead. The run cannot see a verdict recorded after it, so it names both.
• The run.sh header, the carry-forward-wait-seconds input description and the README section are updated to match.

Verification

• bash .github/actions/ci-status/run.test.sh: all cases pass on this branch.
• The same test file run against main's run.sh fails 5 new cases:
  • Behavior: "a stale failure waits for the in-flight re-run of its writer" and "a writer re-run that outlasts the ceiling fails closed" fail on main.
  • Message only: "two contract-only siblings on a failed SHA do not wait on each other", "a missing status still fails after the re-run ends without recording one" and "the failure message names the re-run-this-run remedy" fail on main only on the new message text. main already behaves correctly in the first two. They are regression guards for this change.
• Mutants of this branch's run.sh:
  • A failure held open by any in-flight sibling fails the "two contract-only siblings" case and three others.
  • run_attempt > 1 in place of the writer test fails the "two contract-only siblings" case.
  • Dropping the start-time term fails "the writer attempt that recorded the failure is not waited on".
• The fixture timestamps come from the real runs. The failure status 55241522644 (created_at 05:56:46Z) names run 36673988526 in target_url. That run's attempt 1 run_started_at is 05:33:54Z and attempt 2's is 05:57:10Z. The consumer pin 4610c31e writes the same target_url shape.
• shellcheck on run.sh and run.test.sh, shfmt -d .github/actions/ci-status/, actionlint, markdownlint-cli2 README.md, editorconfig-checker: clean. typos flags only a queueing on run.sh:20, which is already on main. node --test .github/scripts/ci-fanout-consolidation.test.cjs: pass.
• Not verified here: a live run of the new composite. That needs a release and a consumer repin, and claude-code-plugins#4670 tracks it.

Related

• fix(ci): contract-only ci-status reads a stale failure during a re-run claude-code-plugins#4670: consumer tracker and reproduction (throwaway PR claude-code-plugins#5581). It stays open until a release is repinned there.
• feat(ci-status): bounded carry-forward wait and the branched contract-only concurrency group #557, fix(ci-status): order the carry-forward wait by run id so same-second siblings are seen #559, ci-status: carry-forward wait misses a same-second full sibling with a higher run id #561, fix(ci-status): wait on any in-flight sibling before reading ci-lanes #562: earlier carry-forward wait fixes.

🤖 Generated with Claude Code