Milestones
Condensed from docs/IMPLEMENTATION_PLAN.md. Live checkboxes: Progress Tracker.
Goal: every verifiable SPEC in Contracts becomes VERIFIED with real captured responses. No product code.
Key tasks: CPAI probe + fixtures · GPU/contention check · ao_reach hello spike · TalkingHead API note · Pi baseline · audio routing ADR.
Exit: real CPAI JSON for detect/hit/miss · miss-shape answered · CUDA confirmed · reach_hello works · TalkingHead spoke on desktop · ADR 0001 written · BASELINES populated.
Goal: three processes start, parse config, structured JSON logs, loopback WS, clean SIGTERM.
Key: typed config (unknown keys fatal) · logger + Span · local WS 8777/8778 · clients with backoff · systemd units · remote Dev Loop.
Exit: analyze/ruff/mypy clean · 10 min uptime · reconnect <5s · doctor/logs/deploy/rollback · LAN bind refuses without gates.
Goal: Dart CPAI client + frame grabber + poll loop emitting vision events. No state machine yet.
Key: long-lived ffmpeg grabber · detect/recognize · vote-based identity · fps switch · fake CPAI from fixtures.
Exit: offline tests green · live resolve userid <2s with recognize ~once · no fd leak · vision.degraded works.
Goal: CONTRACTS §8 as pure logic with injected clock, 100% branch coverage, all invariants.
Key: sealed events/states · effects as data · effect runner · turn_id · fake-clock timers · dev console.
Exit: property test 10k sequences · golden scenarios · no await in handle() · console walks the ladder.
Goal: wake + VAD on real Pi mic, tuned threshold, PCM to bridge.
Key: capture ring · train hey_comstar.onnx · runtime wake · VAD + pre-roll · streaming · ROC table · do not drop threshold to fix a bad mic.
Exit: 0 false accepts on 60 min room · ≥90% at 3 m quiet · ≥70% with TV · ROC committed · idle CPU <15% one core.
Goal: face recognition opens identity-mapped AO session; typed question → spoken-appropriate text.
Key: session manager · greeter/voice overlays · guest restrictions · terminal + vision MCPs.
Exit: p95 <10s text turn · no markdown in answers · 20 open/close leave zero overlays · guest cannot reach HA.
Goal: speak a question, hear an answer. No avatar required for this gate.
Key: STT/TTS clients · runner wiring · fallback WAVs · latency report · half-duplex barge-in policy · Reach speech prefer (ADR 0003).
Exit: 20 consecutive spoken answers · p50 <8s / p95 <15s · orchestration dominates · fallbacks offline · wake blocked during playback.
Goal: avatar on Pi screen driven by bridge messages.
Key: kiosk shell · SVG/TalkingHead · state visuals · render path ADR · WebGL failure fallback · reconnect.
Exit: ≥24 fps on chosen path · lip-sync acceptable (when GLB) · ADR written · kiosk death does not wedge bridge · voice survives GLB failure.
Goal: the demo. Walk up. Greeted by name. Ask. Spoken answer. Under 15s.
Key: full wiring · greeter <1.5s · follow-up window · stranger modes · enroll script · demo mode overlay.
Exit: 10/10 greet attempts · greeting p95 <2s · turn p95 <15s · guest restricted · stranger can use without instructions.
Goal: survives a week unattended.
Key: failure matrix · backoff everywhere · resource limits + watchdog · 24h soak · retune wake from real audio · runbook · privacy audit.
Exit: 24h zero crashes · ≤2 false wakes · flat memory/fds · auto-recover every failure row · privacy claims verified.