Fix crash when feed syncers ran outside Oban

Feed drops now ingest reliably when no model-backed agents are configured. Previously, the RSS syncer ran as a supervised timer process and could call into the job queue before it was up, which crashed instances without LLM configuration. The syncers and classification work now run through Oban with a cron schedule, so background syncing every fifteen minutes plus boot retries and deduplication are visible and stable. Manual feed syncs from the dashboard enqueue the same worker.