Log the prompt, not just the 500
AI features fail softly: slow, empty, or slightly wrong. We trace provider latency, token counts, and retrieval hit rates next to the user id.
Every AI path has a deterministic fallback. A search result is better than a spinner that never resolves.
