Transcribe an audio file

OpenAI-compatible bounded file transcription. Use Authorization: Bearer <API_KEY> and multipart/form-data. The initial profile supports files up to 25 MB and json or verbose_json responses. Unsupported OpenAI fields are rejected explicitly.

The accepted model set is a curated per-environment allowlist that a model joins only after passing contract qualification; openai/whisper-large-v3 is the launch model. Requesting a model outside the allowlist returns a 400 without dispatching audio.

This endpoint is independently feature-gated during preview rollout. A disabled deployment returns an OpenAI-shaped 404 without dispatching audio. It is synchronous and does not use prediction polling or streaming.