fix(prover-client): don't persist stale proving job results (racing epoch GC)

[AztecBot]

Summary

Follow-up to #22508. That PR kept in-progress jobs from stale epochs alive, but cleanupPass() still calls deleteAllProvingJobsOlderThanEpoch(...), which closes and deletes the per-epoch LMDB store. When an agent then reports an error/result for such a stale in-progress job, the broker's rejected/fulfilled path calls database.setProvingJobError/Result(...), whose batch-queue commit runs batchWrite on the just-closed store — producing Error: Store is closed and failing the new Retries > does not retry if job is stale test.

Skip the DB persist when the job is already stale. The in-memory resultsCache + promise resolution still happen, so callers still observe { status: 'rejected' | 'fulfilled', ... }. There's no point writing to a per-epoch store that is on the way out.

Details

• Flaked 5/10 on merge-train/spartan HEAD locally; passes 20/20 with this change.
• Applied the same guard to #reportProvingJobSuccess to close the symmetric path.
• Full analysis: https://gist.github.com/AztecBot/e093ac749ff8829e127e7e401cd422f6

Test plan

• yarn workspace @aztec/prover-client test src/proving_broker/ — all 7 files pass (145/145).
• Stress-ran proving_broker.test.ts x20 — all green.
• ./bootstrap.sh format / ./bootstrap.sh lint clean.

Failing CI run that motivated this fix: https://github.com/AztecProtocol/aztec-packages/actions/runs/24714302818

ClaudeBox log: https://claudebox.work/s/2bbcd00ade5c0a6a?run=1

[AztecBot]

Automatically closing this stale claudebox draft PR (no updates for 5+ days). Re-open if still needed.