Auto-search warm pool is lost when the search worker idles out after 5 minutes; no-hint decisions aren't logged #773

Update: measured afterwards, the 5-minute worker idle-out is not what suppresses hints (attach rate is the same either side of 5 minutes); most no-hint turns are below the relevance threshold. See this comment.

Magic Context: 0.49.20, OpenCode 2.1.4, macOS, Voyage embeddings, memory disabled. Follow-up to #762 — thanks for the fix.

The warm pool now survives the 1 s deadline, but it lives in the search worker, which is disposed after SEARCH_WORKER_IDLE_MS = 5 * 60_000. With a long cache TTL our turns are often more than 5 minutes apart, so the next turn starts cold, warmOnly skips the chunk lane, and the vector side rarely contributes to hints. A cold turn also pays ~270–350 ms for worker startup.

Evidence is circumstantial: after restart, a turn at 20:28:18 (cold) and the next at 20:33:32 (5 min 14 s later, past the idle timer) both ended empty. I can't prove the pool state because:

• No-hint decisions (empty, too-short, below-threshold) write no log line on the writeNoHintAndReconcile path; only session_meta records them.
• The successful path never logs chunkPool state; the new per-step timing line only appears on a missed deadline.

Asks:

• Log the no-hint reason and pool state (warm/cold, rows loaded) on every decision, at least at debug.
• Keep the decoded pool past the worker's idle timeout (share it outside the worker, or a longer/idle-aware timeout for sessions with a pool), or warm it in the background after a turn so the next one finds it ready.