perf: optimize sync with mtime-based change detection and configurable interval

[troshab]

Summary

Two improvements to the background file sync mechanism:

1. mtime+size based change detection

Instead of reading and SHA-256 hashing every file on each sync cycle, the synchronizer now:

1. stat() each file to get mtime + size (~100x faster than read+hash)
2. Compare with cached mtime/size from previous snapshot
3. Only read + hash files where mtime or size actually changed
4. Reuse cached content hash for unchanged files

Impact for a 17k-file codebase with 0 changes: ~200ms (stat only) vs ~10s (read+hash all).

File stats are persisted in the snapshot alongside content hashes. Old snapshots without stats are handled gracefully (falls back to full hash on first run after upgrade).

2. Configurable sync interval via SYNC_INTERVAL_SECONDS

• Default changed from 300s (5 min) to 60s (1 min)
• Configurable via SYNC_INTERVAL_SECONDS environment variable
• Minimum 10s floor to prevent excessive polling
• The shorter default is practical now that per-cycle cost is dramatically lower

Test plan

• pnpm build compiles without errors
• Backward compatible with old snapshot format (no fileStats field)
• Verify sync detects file changes correctly (add/modify/delete)
• Verify unchanged files are not re-hashed (check logs for hash skip count)
• Verify SYNC_INTERVAL_SECONDS=30 changes the sync interval

Closes #267