fix(sync): clear exhausted jobs blocked by durable read failures

[cursor]

Summary

Stops the sync self-heal sweep from re-logging the same exhausted calendarListSync job every ~10 minutes when the connection already carries a durable provider refusal.

PostHog cluster (7d):

• Sync self-heal sweep exhausted budget — 118 occ, still firing through 2026-08-09 12:52Z for job 6a775383… / connection 6a653974…
• Companion Sync job engine failed — calendarListSync half quiet after attempt 20 at 02:54Z; root cause ProviderCalendarError(discoveryFailed): … The user must be signed up for Google Calendar.

Root cause of remaining noise: #2696 drops durable discovery failures going forward, but a pre-deploy exhausted failed row kept its coalescing key and made self-heal logger.error on every sweep. Events resources on the same connection already had lastReadFailureAt from incrementalPull drops.

Fix: when listing exhausted jobs, auto-clear any whose connection already has a durable lastReadFailureAt marker (same effect as manage-failed-jobs clear), log once at warn, and only error for remaining exhausted rows that still need an operator.

Simplicity

One extra resource lookup on the existing exhausted sample path — no new job states, no new failure classes. Reuses jobs.remove / listByConnection / the health marker already stamped by readFailed and discoveryFailed drops.

Automated validation

• bun test:sync -- packages/sync/src/domain/failed-job-requeue.service.db.test.ts → 7 pass
• bun run lint → no new issues in touched files (pre-existing warnings only)

Independent review

Diff-first pass: auto-clear is gated on an existing durable marker (not on job kind alone), so genuine exhausted transient failures still alert. Cleared jobs free the coalescing key so rediscovery can enqueue a fresh calendarListSync that #2696 will drop. No further highs found.

Test plan

• Exhausted job with lastReadFailureAt on the connection is cleared and removed from error count
• Exhausted job without a durable marker still reports as needing operator attention
• Existing requeue / cooldown / missing-requeuedCount cases still pass

Remaining risk / ops

• After deploy, the next self-heal sweep should clear stuck job 6a775383… and silence 019fdc65-…. Resolve both PostHog issues once quiet.
• Account 6a653974… still cannot list Google calendars — enable Google Calendar or disconnect in Compass so rediscovery can succeed.