Implement phase 2.1 (human detection) and 2.2 (notifications): - Detection scorer computes 0.0-1.0 human likelihood from keystroke timing variance, special key usage, typing speed, command diversity, and session duration - Webhook notifier sends JSON POST to configured endpoints with deduplication, custom headers, and event filtering - RecordingChannel gains an event callback for feeding keystrokes to the scorer without coupling shell and detection packages - Server wires scorer into session lifecycle with periodic updates and threshold-based notification triggers - Web UI shows human score in session tables with highlighting - New config sections: [detection] and [[notify.webhooks]] Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>