feature: add heartbeat loop for periodic health checks

[benvinegar]

What

Adds a periodic heartbeat loop to the control agent so Baudbot becomes proactive instead of purely reactive.

New files

• pi/extensions/heartbeat.ts — pi extension that runs a setTimeout timer, reads ~/.pi/agent/HEARTBEAT.md on each tick, and injects it as a follow-up prompt
• pi/skills/control-agent/HEARTBEAT.md — default checklist deployed to ~/.pi/agent/HEARTBEAT.md

How it works

1. Extension starts a timer on session_start (default: 10 min interval)
2. On each tick, reads HEARTBEAT.md — if empty/missing, skips silently (zero token cost)
3. If content exists, injects it as a followUp message with triggerTurn: true
4. Agent processes the checklist, takes action on anything broken, responds briefly if healthy
5. Timer re-arms after each fire

Default checklist checks

• Agent sessions alive (sentry-agent, orphaned dev-agents)
• Slack bridge responsive
• Email monitor running
• Stale worktrees with no matching active todo
• Stuck todos (in-progress >2h, no dev-agent)

Features

• Error backoff: 2x exponential per consecutive failure, max 1 hour — prevents token burn
• Min interval floor: 2 minutes hard limit
• heartbeat tool: status / pause / resume / trigger / config
• State persisted via appendEntry — run count survives session restarts
• Configurable: HEARTBEAT_INTERVAL_MS, HEARTBEAT_FILE, HEARTBEAT_ENABLED env vars
• Deploy: deploy.sh copies HEARTBEAT.md (always overwrites — admin-managed)

Inspired by

OpenClaw's HEARTBEAT.md + cron service pattern, simplified for our architecture (no separate CronService, no isolated sessions, no job persistence — just a timer + Markdown file).

Docs updated

• AGENTS.md — repo layout
• README.md — new Heartbeat section + architecture diagram
• CONFIGURATION.md — new env vars
• Control agent skill — heartbeat section + startup checklist