The deployed-PR-still-queued check gives up after an hour of reads #5014

🤖 Generated by the Agentic Engineer (interactive Claude Code session)

Evidence

Found by the merge-queue failure count in #4993.

• 4 failed merge-group runs between 2026-10-01 and 2026-10-05 failed in job 🧾 Check the deployed PR is still queued, step 🧾 Read the PR's merge-queue state, which gave up after 120 reads (60 minutes).
• 3 of the 4 pull requests later merged with no new push.
• Evidence basis: check annotations.
• Examples: 36799676044, 37311681456, 37368240379.

Problem

These runs are the most expensive per failure in the window (about 513 run-minutes for 4 runs): each holds the queue for an hour and then evicts a candidate that was fine.

Expected behaviour

The check reaches a verdict about the pull request's queue state well within the queue's own timeout, and a read it cannot interpret names what it saw.

Acceptance criteria

• Find which state the 120 reads kept seeing, from the runs above.
• Fix the check so that state produces a verdict, or a failure naming it, without waiting the full hour.
• A test covers that state.

Size: small. Part of #4979.