perf: cache all prepared statements at load time

[mageroni]

Summary

db.prepare() was called inline on every poll cycle (every 10s) and on every tool invocation, causing repeated SQL compilation overhead that was completely avoidable.

Change

Pre-compile all statements once into a stmts object at module load time. Every db.prepare(...).run/get/all(...) call is replaced with stmts.<name>.run/get/all(...).

No behaviour changes — only the compilation cost is eliminated.

Impact

Over an 8-hour session with 10s polling:

• ~2,880 poll cycles × 5+ prepare() calls each = ~14,400+ unnecessary compilations avoided
• Tool invocations (send_message, reply_to_message, get_message, get_agents) each eliminate 1–4 prepare() calls

Bonus: get_agents now uses SELECT with explicit columns instead of SELECT *, avoiding fetching the unused metadata and registered_at columns.

Part of a larger perf audit

This is the highest-impact / lowest-effort item from a 10-item performance audit:

1. ✅ Cache prepared statements ← this PR
2. Run cleanupStaleData periodically (not just at startup)
3. Add composite index for rate-limit query (sender, recipient, created_at)
4. Add index on created_at for cleanup queries
5. Add PRAGMA mmap_size for memory-mapped I/O
6. Batch rate-limit checks across the message poll batch
7. Deduplicate git-walk logic in onSessionStart
8. Skip queue depth COUNT when message count is low
9. Skip re-registration in onSessionStart when cwd is unchanged
10. Prune agent_description from getActivePeers when not displayed