Overseer voice: wire estate Chatterbox TTS behind VoiceTransportShim (kill ElevenLabs readback dependency)

[heavygee]

Goal

Kill the ElevenLabs dependency for overseer voice readback by wiring the estate-local Chatterbox TTS (already operated on proxmox, OpenAI-compatible HTTP) behind the existing VoiceTransportShim. Free + local, minimal engineering. This is a bridge win that is independent of the larger agentic-overseer plan — worth doing regardless of where the full-duplex / brain decisions land.

Context

• Overseer audio summaries in headset work today via ElevenLabs TTS (AGENT_NOTIFY_SUMMARY → corruption-tolerant parse → event/inbox → TTS readback). Operator does not intend to keep paying ElevenLabs.
• Evaluation + framing: docs/plans/2026-07-29-unmute-voice-evaluation.md (committed). Destination is a free/local agentic voice overseer; mouth/ears are nearly solved on the estate (Speaches STT + Chatterbox TTS). This ticket is Phase 0/1 (free local readback).
• Existing seam: feat/overseer-stt-tts-endpoints — shared/src/voiceTransport.ts (capabilities {stt,tts}, models, default voice) + VoiceTransportShim { transcribe(), synthesize() } + registry in hub/src/voice/transport/. Chatterbox is OpenAI /v1/audio/speech-shaped → drops in with almost no protocol invention.

Scope

• Add a local-openai (or chatterbox) entry to the voice transport registry (hub/src/voice/transport/).
• Implement synthesize(text → audio) against Chatterbox /v1/audio/speech (base URL + voice id config).
• Wire config so overseer readback can select this backend (env / settings).
• Voice-id mapping + a sane default voice.
• Tests with fixture audio (mirror existing transport tests).
• Dogfood: point one overseer summary readback at Chatterbox on :3006, confirm parity with ElevenLabs on the same summary strings.
• (Optional follow-up, separate) STT via Speaches for PTT dictation — Phase 2, not this ticket.

Out of scope

• Full-duplex conversation / VAD / barge-in (that's the agentic Phase 3, X-vs-Y decision).
• The local tool-using LLM "brain" (separate spike).
• Kyutai / unmute (upgrade path only if Chatterbox fails quality/latency dogfood).

Done when

ElevenLabs can be switched off for overseer summary readback with Chatterbox covering it, free + local, dogfooded on :3006.

[heavygee]

Retarget from #29 (2026-08-06)

Peer decision: keep the June VoiceTransportShim substrate, discard the cloud-endpoint chase.

Plan (next commits)

1. Fresh branch off upstream/main (e.g. feat/overseer-chatterbox-tts) — not a 302-commit rebase of feat/overseer-stt-tts-endpoints.
2. Port only: shared/src/voiceTransport.ts (types + capabilities), hub/src/voice/transport/{types,index}.ts, tests skeleton. Extend registry for local-openai / chatterbox.
3. Implement synthesize → estate Chatterbox OpenAI-compatible POST /v1/audio/speech (base URL + voice from env).
4. Wire overseer readback selection (env first; settings later if needed). Keep ElevenLabs synthesize as optional A/B for dogfood.
5. Drop / do not port: Gemini/Qwen STT shims and any STT route that duplicates upstream /voice/transcription.
6. Dogfood on :3006 after operator-approved soup — not promoting the stale Voice provider abstraction: expose standalone STT + TTS endpoints #29 branch.

No PR until operator says go. Does not block A2A P1 (tiann#1374) or notify-settings (tiann#1375).