feat(jobs,extraction): process requests in the worker with retries and visible errors #7

Goal

The worker drains processing jobs, calls the AI service, persists run, fields and evidence and moves the request to REVIEW in one transaction; failures retry with backoff and end in ERROR with a visible cause.

Acceptance criteria

• pg-boss 12 with transactional enqueue (fromDrizzle), singletonKey = requestId, retry with backoff, dead-letter queue
• The handler is idempotent: delivering the same job twice yields exactly one extraction run
• AI-service timeout or 5xx → retry; after the retry limit → ERROR with cause and attempts shown in the request list; reprocess re-enqueues
• drain({ maxMs }) processes jobs until its budget; the worker loop uses it
• Logs carry requestId, jobId, companyId and no document content

Not part of this task

• Vercel triggers for drain() (showcase epic), export

Affected areas

• src/features/jobs/
• src/features/extraction/
• src/features/requests/

Test plan

Security/Privacy affected?

Job payloads carry IDs only; handlers run inside withTenant.

Epic: #2 · Architecture: docs/decisions/ADR-0001-pilot-architecture.md