Fix duplicate ingestion jobs starving newly enabled sources

[ralyodio]

The production ingestion queue accumulated over 74,000 waiting jobs because each scheduler tick created a new minute-based job ID for sources that were already queued. Those sources also occupied the first page of the due-source query, preventing 38 newly enabled California police feeds from reaching a worker.

Use BullMQ deduplication per source for scheduled and manual runs, lasting through waiting and execution. Page past sources already represented in Redis so later overdue sources can be admitted, with at most 50 new requests per tick. Unique job IDs retain completed history without blocking future runs.

Adds an explicit queue repair CLI: audit by default; --apply backs up pending metadata and creates or locates a deduplicated replacement before removing legacy requests. It preserves force flags and leaves active jobs, unrelated job types, source configuration and ingested records untouched. Apply after deployment to consolidate the existing backlog.

Validation: 1,641 local tests passed, including four real-Redis integration tests for concurrent deduplication, rescheduling after success/failure, pagination past queued sources and safe queue repair; one local PostgreSQL migration test skipped. Eight discovery checks passed, lint passed with existing warnings. CI now provides Redis for these integration tests. Separately, initial production reads succeeded for all 39 California police sources, loading 230 announcements.