ElevenLabs Voice Design (/v1/text-to-voice) record + replay #452

Summary

aimock records and replays ElevenLabs text-to-speech (onElevenLabsTTS), sound generation and music, but not Voice Design (/v1/text-to-voice/*). An app that designs a voice before speaking with it can't run hermetically: the design call either misses (strict mode fails) or goes to the real ElevenLabs account and bills it.

Checked against v1.42.0.

Routes

Example design body (@elevenlabs/elevenlabs-js):

{
  "voice_description": "A weathered sea captain in his sixties, gravelly, unhurried",
  "model_id": "eleven_ttv_v3",
  "auto_generate_text": true,
  "should_enhance": true,
  "guidance_scale": 25,
  "quality": 0.5
}

Suggested behaviour

• Match design fixtures on voice_description (string or RegExp), like onElevenLabsTTS matches on text, e.g. mock.onElevenLabsVoiceDesign(/sea captain/, { previews: [...] }).
• Save: replay should accept a generated_voice_id that a design fixture returned and hand back a stable voice_id, so a following /v1/text-to-speech/{voice_id} call can match an existing TTS fixture. It's fine if save isn't matched separately and just echoes a deterministic id.
• Record: proxyAndRecord under the existing elevenlabs provider key. Worth noting: each preview's audio_base_64 is a full MP3 (usually 3 per call), so these fixtures will be much larger than JSON-only ones.
• Nice to have: GET /v1/voices/{voice_id} and DELETE /v1/voices/{voice_id}. Apps that manage slots call them after saving (delete should be idempotent on 404).

Workaround today

mock.mount('/v1/text-to-voice', handler) with a hand-written Mountable. It works, but it bypasses the recorder and journal, so there's no record path.