[squid-protocol]

Part of #4840

Two quiet failures from 2026-10-10. Each had a fix in place on the spot, but nothing would have made either one visible the next time. This makes both visible.

1. A PR that lost its approval now carries a label: shepherd:reapprove

What happened: #4858 was green but didn't merge. A push dropped shepherd:merge (by design: an approval is bound to the head it was given on). With no request label left, the bot also cleared every status label, so the only trace was a comment. Nobody re-added the label until a person noticed.

Now: when the bot drops shepherd:merge because the head moved, it also sets shepherd:reapprove.

• It stays after the request labels are gone, so is:open label:shepherd:reapprove lists every PR waiting for a maintainer to re-approve.
• It clears when shepherd:merge comes back (or shepherd:hold is set), or when someone removes it by hand. A manual removal sticks.
• Reconcile already visits PRs that carry any status label, so these stay in its sweep.
• The security model is unchanged: nothing is re-approved automatically.
• The "awaiting marker" case (push or label time unknown) stays quiet as before. That state is undecided and can resolve on the next pass, so a label there would flap.

2. A hung test now fails by name instead of hitting the job time limit

What happened: every macOS matrix leg on main ran into the 45-minute timeout-minutes on a hung test (#4862: no flock on macOS). GitHub reports that as "cancelled", which names nothing, and it read as noise all day.

Now:

• full-suite-gate.yml and smoke-test.yml run pytest with --timeout=600 (pytest-timeout==2.4.0). The slowest real test is ~60 s (.test_durations), so this is ~10× headroom and well under the 30 and 45-minute job limits. Checked locally with pytest 9.1.1 + xdist 3.8.0: a hung test fails as FAILED …::test_hangs - Failed: Timeout, and the rest of the run carries on.
• main_watch.py gives each cancelled leg its run time ("… after 45 min"). It also no longer calls a cancel at the time limit "not a verdict on the code".

Also

docs/ci.md: the new label in the status table and its lifecycle, the timeout, and a pair of paragraphs that appeared twice is removed.

Checks

• test_shepherd_gh.py + test_main_watch.py: 69 pass, including 4 new tests.
• pr_gates --fast PASS. --ratchets all PASS except ports (the local CardDemo corpus is not at its pinned ref; same on main).
• Both workflows parse as YAML. actionlint is not installed here.
• main_watch.py reconcile --dry-run against the live repo: no actions. That's correct, since the latest verdict is already on main is red: Full Suite Gate (All OS x Python) #4840.

After merge

• Run shepherd-reconcile.yml once with ensure_labels: true, so the label exists with its colour and description. Until then, the first add creates it with GitHub's default colour.

Not done here (worth a look)

• The macOS legs were split into 3 shards on 2026-10-10 because they "timed out". The real cause was the hang: the shards together take ~18 min, so one macOS leg per Python would fit in 45 min again and save the extra setup time per shard on the costliest runners.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AUWeQX7v4QBdErFbghEEEX

[squid-protocol]

Note for merging: this changes shepherd_gh.py itself, so per the shepherd's own rule it merges through the local shepherd (python tests/tools/box/ci_shepherd.py add 4865) or a manual gh pr merge --squash after review, not by its own label path.

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab