Poll a managed-voice generation started by POST /v1/media/generate-audio.

Returns the audio_generation task envelope. While it is queued or running, poll again no sooner than its retry_after_ms (a 429 here is a delay, never a failure). succeeded carries results (the audio file_ id, and for text to speech the alignment sidecar); failed carries a typed error whose details.spend says whether provider spend may have occurred. Moda is never charged twice and never charges a failed or canceled task.

This route never calls the voice provider and is never plan-gated: it observes work an entitled call already admitted, and re-drives any that stalled. Only the key's own user sees a task; anything else is an identical 404.