Rebuild Reachy voice and motion pipeline

[asmolebot]

What changed

• replace the OpenAI Realtime second-agent architecture with local STT, OpenClaw as the sole conversational brain, and pluggable Chatterbox or ElevenLabs TTS
• add remote Reachy camera, face tracking, sound-direction tracking, playback gain, echo suppression, and stage latency logging
• map the full 81-emotion catalog plus the official dance library with deterministic natural-language selection
• add an always-hot local voice service and focused contract, conversation, playback, provider, and motion tests

Why

The previous Realtime path duplicated the agent, constrained voice selection, exposed only a tiny hard-coded motion set, and allowed tracking to fight authored choreography. This rebuild keeps reasoning in OpenClaw while making speech and movement independently testable and locally deployable.

Impact

Reachy can now use local speech services, preserve a dedicated OpenClaw body session, answer visual questions from live camera frames, track faces or sound, and execute expressive motion without competing control loops.

Validation

• 30 body tests pass
• 2 voice-service contract tests pass
• git diff --check passes
• live Reachy service validation completed against Reachy Mini 1.9.0