pan close refuses landed, deployed issues: deploy probe times out on a stalled event loop; stale 'running' verification artifact blocks row 3 #4543

Problem

pan close refuses landed, deployed issues for two reasons that are not real Definition-of-Done failures. Observed 2026-10-04 09:10-09:16 UTC, closing out 18 issues that were merged and deployed in build cd27acfd:

1. Deploy row (8) times out on a stalled event loop. 6 of 22 close-out attempts failed with dod:deploy — dashboard not reachable at http://127.0.0.1:3011 … The operation was aborted due to timeout: PAN-4509, PAN-4506, PAN-4383, PAN-4514 and PAN-4541 on the first pass. Each succeeded on retry. /api/health normally answers in under 1 ms, but a 0.5 s sampler during close-outs recorded 4.66 s at 09:15:17 and 4.05 s at 09:15:51, each during a running pan close. Failures followed a successful close-out of the previous issue.
2. Verification row (3) blocks on a stale running artifact. PAN-4498 and PAN-4508 fail with dod:verification — verification artifact: running at 2026-10-04T05:44:24Z (and 05:45:37Z), although both PRs passed verification (passed at heads 3fe790ef and f91b6577) and merged with green main CI. Their workspaces/feature-pan-<n>/.overdeck/verification-latest.json is {"status":"running", "head8": null}. It was left by a duplicate verification run (started by a deacon-lite re-request at 05:06) that ended {"outcome":"skipped","reason":"The pull request already merged; pre-merge verification no longer applies."} without updating the latest artifact.

Root cause

1. src/lib/lifecycle/dod-gate.ts line 278 probes once with a 3 s budget:

   const response = await fetch(url, { signal: AbortSignal.timeout(3000) });

   No retry. Separately, something in the dashboard server's close-out handling blocks its event loop for about 4-5 s.
2. checkVerificationRow (dod-gate.ts ~lines 445-468) treats a missing artifact on landed work with green main CI as a pass (!artifact && settlement?.landedWork && settlement.mainVerifyStatus === 'pass'), but a non-terminal running artifact as a miss. The merged-skip path in src/lib/cloister/verification-runner.ts (skipMergedVerification) returns skipped without rewriting the latest artifact to a terminal state.

Fix

1. Deploy probe: retry /api/health up to 3 times with a 5 s timeout per attempt and a 2 s pause between attempts, using fake timers in tests. Report the miss only when all attempts fail.
2. Event-loop stall (implementation checkpoint): capture a CPU profile (node --cpu-prof on the dashboard, or --inspect plus a sampled profile) while running pan close on an issue. Name the blocking call with file:line in the PR. If it is a sync child process or sync file scan on the request path, make it async (repo rule: no execSync in server-reachable code). If no single cause is found within the issue's scope, document the profile findings in the PR and file a follow-up. Do not leave the stall unmentioned.
3. Verification row: treat a non-terminal artifact (running, or any status that is not passed/failed) the same as a missing one when settlement.landedWork and settlement.mainVerifyStatus === 'pass' hold, and say so in the row's observed text. Also make the merged-skip path in verification-runner.ts write a terminal skipped latest artifact, so no new stale running files appear.

Acceptance criteria

• Given readJson fails twice with a timeout and then returns a health body, checkDeployRow passes (unit test, fake timers).
• Given a running verification artifact, landedWork: true and mainVerifyStatus: 'pass', checkVerificationRow returns pass with an observed note naming the stale artifact (unit test).
• Given a running artifact and mainVerifyStatus not pass, the row is still a miss (unit test).
• Given the runner's merged-skip path, the latest verification artifact ends with a terminal skipped status (runner unit test).
• The PR names the event-loop stall's cause with file:line and fixes it, or documents the profile and links a follow-up issue.
• docs/MERGE-WORKFLOW.md (close-out section) describes the deploy-row retry and the stale-artifact rule.
• After merge and deploy, pan close PAN-4498 and pan close PAN-4508 complete with no --accept-* override.