fix(prWorkItems): preserve workItemId across racing pipeline writers

[zbigniewsobiecki]

Summary

PRs flagged with auto on Linear weren't auto-merging. Trace: every recent llmist PR has pr_work_items.work_item_id IS NULL, so pr-ready-to-merge bails with "No work item linked to PR" even though the implementation correctly captured MNG-93 at run time.

Root cause is a three-way race against linkPRToWorkItem between the implementation pipeline and the review pipeline that's queued as soon as the PR opens. The review pipeline captures workItemId synchronously at webhook arrival time (DB-only resolveWorkItemId), gets null because the implementation hasn't yet linked the PR, then writes that stale null back to pr_work_items post-execution — clobbering the correct work_item_id the implementation wrote in between.

Fix — three intersecting changes

1. linkPRToWorkItem Step 2 ON CONFLICT: work_item_id is now COALESCE(EXCLUDED.work_item_id, pr_work_items.work_item_id) — one-way set. A later writer with null cannot erase a known link.
2. linkPRToWorkItem Step 1 prologue: deletes any racing orphan (projectId, NULL workItemId, prNumber) row before promoting the work-item-only row, so the partial-unique-index uq_pr_work_items_project_pr can't fire (the implementation otherwise hits this and silently logs a warn).
3. runAgentExecutionPipeline: re-resolves workItemId via the existing resolveWorkItemId helper at run time and patches agentInput so the corrected value flows into runAgent (and into agent_runs.work_item_id), not just into the post-execution link.

Tests

• tests/integration/db/prWorkItemsRepository.test.ts (+2): preserves an existing non-null workItemId when called with workItemId=null, removes a stale orphan ... row when promoting a work-item-only row. Both fail before the fix.
• tests/unit/triggers/shared/agent-execution.test.ts (+3): re-resolves workItemId from DB when result.workItemId is undefined, preserves trigger-supplied workItemId when DB lookup is unnecessary, leaves workItemId undefined when neither trigger nor DB has one.
• tests/unit/db/repositories/prWorkItemsRepository.test.ts (+2): deletes any racing orphan (NULL workItemId, prNumber) row before Step 1 UPDATE, does NOT call delete when workItemId is null. Updated existing onConflictDoUpdate assertion since set.workItemId is now an SQL expression (COALESCE).

Counts: unit 7814/7814 pass, integration 524/524 pass, lint clean, typecheck clean.

Drive-by cleanups

To leave lint at zero warnings:

• Drop unused renderManifestStep import in web/src/components/projects/pm-wizard.tsx (left over from PR feat(006/5): cleanup legacy — spec 006 complete #1129's manifest migration).
• Rename a pm-conformance describe string '/${id}/webhook convention' → '/<id>/webhook convention' so biome's noTemplateCurlyInString stops flagging it.

Out of scope

• ❌ Backfilling the 15 broken historical llmist rows — separate manual SQL after this lands.
• ❌ Defensive PR-body-parsing fallback in resolveWorkItemId (Linear regex) — explicitly excluded; the real fixes above are sufficient.
• ❌ Refactoring how PROpenedTrigger captures workItemId — Part C handles the consequence at the right layer; rewriting capture would balloon scope.

Test plan

• npm test (7814/7814 pass)
• npm run test:integration (524/524 pass, ~5.5 min)
• npm run typecheck clean
• npm run lint clean (was 2 pre-existing warnings; now 0)
• After merge: kick a Linear-triggered llmist PR, watch pr_work_items.work_item_id stay set through implementation + review, confirm pr-ready-to-merge fires when the reviewer approves.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!