feat(voice): add Voxtral TTS provider (local + cloud)

[RollandMELET]

Summary

• Adds Voxtral (Mistral AI) as a new TTS provider with two modes:
  • Local: mlx-audio server on Apple Silicon (zero cost, ~2.5GB VRAM, high-quality French/English voices)
  • Cloud: Mistral API, OpenAI-compatible endpoint ($16/M chars)
• Adds toOggOpus() helper that converts any audio format to OGG Opus via ffmpeg, with magic-bytes detection to skip if already OGG. This also fixes ElevenLabs (returns MP3) and Gradium (returns raw Opus) which weren't being converted before.
• Updates the TTS cascade to: Voxtral local -> Voxtral API -> ElevenLabs -> Gradium -> macOS say
• Documents all config vars in .env.example

Motivation

Running TTS locally eliminates API costs and latency. Voxtral 4B is Mistral's open-weight TTS model with native French support - ideal for ClaudeClaw users who self-host on Apple Silicon Macs. The cloud fallback uses the same Mistral API for users without local hardware.

Configuration

# Local (Apple Silicon)
pip install mlx-audio
mlx_audio.server --port 8881
# .env
VOXTRAL_LOCAL_URL=http://localhost:8881

# Cloud
# .env
MISTRAL_API_KEY=your-key

# Optional
VOXTRAL_VOICE=fr_male          # or fr_female, en_male, en_female
VOXTRAL_LOCAL_MODEL=mlx-community/Voxtral-4B-TTS-2603-mlx-4bit

Test plan

• Local mode: set VOXTRAL_LOCAL_URL, send a voice note, verify OGG Opus response
• Cloud mode: set MISTRAL_API_KEY only, verify cloud API fallback works
• Cascade: disable Voxtral vars, verify ElevenLabs/Gradium/macOS say still work
• toOggOpus(): verify MP3 (ElevenLabs) and raw Opus (Gradium) are properly converted
• No Voxtral config: verify existing behavior is unchanged

Breaking changes

None. Fully backwards-compatible - existing setups without Voxtral vars work exactly as before.