[TheGreatAxios]

Closes CL-11521

Why

The manager evaluated a judged head a second time by replaying the judge's answers. When an answer had been dropped, or the pull request had changed in the meantime, the comment said "Some rules could not run this time" and the pull request got Needs Human Review. Both the judge's result and the code review's were forgotten once nothing was waiting, so the next event on the same head started each of them again.

What changed

• The judge reports the whole verdict it evaluated: the comment, labels, Dos and rule results.
• When the manager has the judge's verdict for the current head and this pass's rules still need the judge, it posts and executes that verdict as reported. It does not evaluate the head again. Before posting, it checks that the verdict names this pull request, head and host.
• If this pass's rules no longer need the judge (a failed hard rule, a draft or a closed pull request), the pass's own verdict goes up instead.
• While CI or other host checks are still running, the manager does not start the judge. It marks the head waiting, as it does for a policy without the judge, and starts the judge on the first pass after the checks finish.
• A decided verdict stays on its head. "Decided" means it does not wait, is not degraded, has every check confirmed and holds no Do for checks still running. A later event on the same head posts that verdict again and starts no second judge.
• The manager does not keep an undecided or degraded verdict, a failure, a missed deadline, or any result over 256 KiB. That keeps done's body under 1 MiB. Within the pass that has the result, the judge is not started again. The next event starts it again.
• A finished code review stays on its head the same way. A later event on that head posts the stored review with the verdict, but only while this pass's rules still let code review through. It does not start a second review.
• The manager no longer replays the judge's answers, and the verdict detail the portal reads no longer carries them.
• Docs: the ARCHITECTURE.md judge and manager bullets, plus the triage-manager and triage-judge READMEs.
• User manual: unchanged. It already describes the Waiting for CI label and the verdict comment, and a person still sees the same labels and comments. The only visible difference is that the wrong "could not run" comment stops appearing.

Verified

• bun --conditions=intx-src test packages/triage-manager packages/triage-judge packages/triage-contracts packages/triage-workflows/src: all pass.
• The new and changed manager tests fail without this change. They check that:
  • the posted verdict is exactly the judge's, and later events on the same head post it again without a second judge;
  • a draft wins over a stored verdict;
  • a verdict naming another pull request fails the pass;
  • an oversized result is not stored;
  • while CI runs the judge waits and the head is marked waiting, and after CI passes an undecided stored verdict is judged again and its Dos run;
  • a failed or late judge or review is not kept;
  • a stored review is posted again, and is not posted once code review is off.
• The only scoring change is a waiting branch in renderVerdict for items marked judgeSkipped, which only the manager sets. The offline triage-workflows evals pass (2), and the paid-model eval was not run.
• The hub's manager route and state Postgres tests pass (27).
• bun run check, bun run lint and the policy check against origin/main are clean.

[linear-code]

CL-11521

[corbits-triage]

Triage: Unsure

Waiting on:

• checks are still running: check

Some rules could not run this time. Triage will re-check this pull request.