perf: fix stale startlock delay and add comprehensive benchmarks (#484)

* fix(daemon): check for stale startlock before waiting 5 seconds

When a previous daemon startup left behind a bd.sock.startlock file
(e.g., from a crashed process), the code was waiting 5 seconds before
checking if the lock was stale. This caused unnecessary delays on
every bd command when the daemon wasn't running.

Now checks if the PID in the startlock file is alive BEFORE waiting.
If the PID is dead or unreadable, the stale lock is cleaned up
immediately and lock acquisition is retried.

Fixes ~5s delay when startlock file exists from crashed process.

* perf: add benchmarks for large descriptions, bulk operations, and sync merge

Added three new performance benchmarks to identify bottlenecks in common operations:

1. BenchmarkLargeDescription - Tests handling of 100KB+ issue descriptions
   - Measures string allocation/parsing overhead
   - Result: 3.3ms/op, 874KB/op allocation

2. BenchmarkBulkCloseIssues - Tests closing 100 issues sequentially
   - Measures batch write performance
   - Result: 1.9s total, shows write amplification

3. BenchmarkSyncMerge - Tests JSONL merge cycle with creates/updates
   - Simulates real sync operations (10 creates + 10 updates per iteration)
   - Result: 29ms/op, identifies sync bottlenecks

Added BENCHMARKS.md documentation describing:
- How to run benchmarks with various options
- All available benchmark categories
- Performance targets on M2 Pro hardware
- Dataset caching strategy
- CPU profiling integration
- Optimization workflow

This completes performance testing coverage for previously unmeasured scenarios.

* docs: clarify daemon lock acquisition logic in comments

Improve comments to clarify that acquireStartLock does both:
1. Immediately check for stale locks from crashed processes (avoids 5s delay)
2. If PID is alive, properly wait for legitimate daemon startup (5s timeout)

No code changes - only clarified comment documentation for maintainability.

---------

Co-authored-by: Steve Yegge <steve.yegge@gmail.com>

cmd/bd/daemon_autostart.go

@@ -183,7 +183,18 @@ func acquireStartLock(lockPath, socketPath string) bool {
 	// nolint:gosec // G304: lockPath is derived from secure beads directory
 	lockFile, err := os.OpenFile(lockPath, os.O_CREATE|os.O_EXCL|os.O_WRONLY, 0600)
 	if err != nil {
-		debugLog("another process is starting daemon, waiting for readiness")
+		// Lock file exists - check if it's from a dead process (stale) or alive daemon
+		lockPID, pidErr := readPIDFromFile(lockPath)
+		if pidErr != nil || !isPIDAlive(lockPID) {
+			// Stale lock from crashed process - clean up immediately (avoids 5s wait)
+			debugLog("startlock is stale (PID %d dead or unreadable), cleaning up", lockPID)
+			_ = os.Remove(lockPath)
+			// Retry lock acquisition after cleanup
+			return acquireStartLock(lockPath, socketPath)
+		}
+
+		// PID is alive - daemon is legitimately starting, wait for socket to be ready
+		debugLog("another process (PID %d) is starting daemon, waiting for readiness", lockPID)
 		if waitForSocketReadiness(socketPath, 5*time.Second) {
 			return true
 		}