[Bug]: /v1/audio/speech with stream=true + response_format=pcm concatenates multiple TTS generations into one stream (~2× audio_duration vs WAV)