feat: rhythm engine stage 1 — observe-only day-phase inference

[cnewkirk]

Summary

Stage 1 of the Rhythm Engine: the daemon infers the occupant's day phase (dawn → morning → daylight → evening → wind_down → night → sleep), detects sleep onset and wake, learns wake/bed anchors — and drives nothing. Observe-only: zero bridge writes (enforced by a test client that asserts on any write), zero behavior change when rhythm: is absent from config.

• rhythm: config block — bed target, weekend drift cap, presence tunables, signal-file paths, rollout stage (only "observe" starts; mornings/full refuse with a ConfigError until those stages ship).
• presence.py — per-room activity judging with pet discounting (light change = human; room-to-room progression = human; motion near a light change = human; solo single-room motion = pet). Every judgment carries its rule name.
• rhythm_control.py — learned-anchor store (bounded to 14 samples per weekday/weekend class, corrupt-file-safe JSON) + the phase engine (anchored, negotiated transitions; sleep-onset confidence vote; sustained-wake confirmation; night wakings explicitly rejected from the wake anchor; restart-safe seeding verified by a 24-hour sweep test).
• Daemon plumbing — MotionAware SSE routing, .next-alarm/.phone-charging signal files (same control-file channel as TV bias), phase changes logged at INFO with the full evidence dict, atomic state persistence (tmp + rename), rhythm failures isolated from the drive loop, all rhythm state lock-guarded on both threads.
• hueman rhythm — status command: phase, target vs learned anchors, weekly phase error in minutes, last decision's evidence. Degrades cleanly on malformed state files; needs no bridge access.
• Docs: README section, CLAUDE.md pointer, example config block.

Process & verification

Built task-by-task with a fresh reviewer gating each task, then a whole-branch final review, which caught and drove fixes for two restart-path Criticals (morning restart stranding the engine in night; cold-start recording a fabricated sleep onset) and a design hole where nightly bathroom trips would poison the learned wake anchor. All fixes verified by direct simulation against the shipped head.

• 356 tests passing (49 new), every behavior change RED-first
• mypy --strict: clean
• Homelab-identifier sweep over the full diff: clean

After merge (ops side)

Deploy with stage: "observe" and log_level: debug (night-waking rejections trace at DEBUG), let it run ≥5 days, compare hueman rhythm + the rhythm: log lines against reality before designing stage 2 (dawn ramp) on the observed data.

🤖 Generated with Claude Code