[somaz94]

Non-streamed WAV ("stream": false) kept the streaming placeholder 0xFFFFFFFF in the RIFF and data chunk sizes, so readers that trust the header, like Python's wave, saw a one-sentence file as about 25 hours long. StreamingAudioWriter reads the buffer before the container closes on purpose, since a stream cannot know its final length, and generate_audio joined those chunks as they were. Once the chunks are joined the length is known, so generate_audio now writes the real RIFF size and walks the chunks to data (ffmpeg puts a LIST chunk before it, so it is not at offset 36) to write its size too. Streaming responses are unchanged. The non-stream branch of /v1/audio/speech, its download link and /dev/captioned_speech all go through generate_audio, so all three get the fix.

The new test feeds real StreamingAudioWriter WAV chunks through generate_audio and checks the RIFF size and wave.getnframes(). It failed before the change with a RIFF size of 4294967295. uv run pytest gives 669 passed and 9 skipped, ruff check . passes, and the changed files are formatted. I used Claude Code to help trace this and write the change, and I reviewed and tested it myself.

Closes #586

Testing

• uv run pytest passes locally (API changes)
• npm run test:web passes locally (web player changes)
• New or changed behavior has a test
• Ran in Docker (npm run cpu:up / gpu:up), for inference or docker changes
• CHANGELOG.md entry, if users will notice the change

[remsky]

Thanks for the fix, merged in @somaz94