Fix Ingest Hangs And Duplicate Haiku Source Fetches

The haiku API call in agent-luce’s ingest phase is hanging for hours, triggering systemd timeouts with negligible CPU use - proof it’s stuck waiting on a single network response. Meanwhile, the task-loop fetches critical sources twice: once from built-in blocks and again from the legacy PROJECTS.yaml, wasting bandwidth and risking stale data. This duplication isn’t just wasteful - it’s a blind spot in reliability.

Here’s the breakdown:

• The ingest step times out at ~4.75 seconds over 80 seconds, signaling a stuck call.
• Logs repeatedly show ‘Fetching source: email’ and ‘Fetching source: github’ twice.

Culturally, this mirrors a broader pattern: automation teams prioritize speed but overlook source redundancy until failures cascade. The real fix is twofold: first, tighten ingest with timeout 300 on claude to avoid systemlock; second, clean up the PROJECTS.yaml loop - either remove it or deduplicate sources. Don’t assume built-in is foolproof; verify it’s the only source. All sources should be fetched once, not duplicated like a tech echo chamber. When does your pipeline stop waiting on the same response - and start trusting each source just once?