fix(tts): transcode xAI ogg requests to Opus

[ishaanlabs-gg]

Fixes #57213.

Summary

• synthesize xAI TTS .ogg requests to a temporary MP3 because xAI does not provide native Opus/OGG output
• transcode that temporary MP3 to the requested .ogg target so Telegram auto voice replies receive real Opus audio
• cover the provider path with a regression test that verifies .ogg requests do not keep MP3 bytes under the .ogg name

Tests

• RED: scripts/run_tests.sh tests/tools/test_tts_xai_speech_tags.py -q failed with out.read_bytes() == b"mp3" before the fix
• scripts/run_tests.sh tests/tools/test_tts_xai_speech_tags.py -q
• scripts/run_tests.sh tests/tools/test_tts_xai_speech_tags.py tests/tools/test_tts_opus_routing.py tests/gateway/test_auto_voice_reply_format.py -q
• .venv/bin/python -m py_compile tools/tts_tool.py tests/tools/test_tts_xai_speech_tags.py
• git diff --check

[alt-glitch]

This was generated by AI during triage.

Related: implementing fix for #57213 (xAI TTS writes MP3 bytes into .ogg voice files). Part of the broader .ogg/Opus TTS cluster — sibling of #57190 (Edge branch, #57048) and the OpenAI-compatible fixes #54597 (earliest canonical transcode)/#56873. Same symptom class, different provider branch in tools/tts_tool.py. Not a duplicate.

[teknium1]

Thanks for isolating the xAI provider path. The reported behavior is still present on current main: Telegram auto voice replies request .ogg in gateway/run.py:13300-13311, while xAI writes MP3 response bytes to the requested output path in tools/tts_tool.py:1370-1413.

Problems

• tools/tts_tool.py:1288 raises when _convert_to_opus() returns no path, before the temporary MP3 cleanup at :1292-1295. _convert_to_opus() returns None without ffmpeg on current main (tools/tts_tool.py:917-918), leaving the UUID-named MP3 behind. Gateway finalization only removes the requested/returned paths (gateway/run.py:13357-13362).

Suggested changes

• Move intermediate-file cleanup into a finally block and add a regression test for the failed-conversion path.

Automated hermes-sweeper review.

[teknium1]

When _convert_to_opus() returns None (including when ffmpeg is unavailable), this raises before the cleanup below, leaving the UUID-named temporary MP3 behind. Please clean up api_output_path in a finally block and add a failure-path regression test.

[teknium1]

Resolved at the class level by PR #73072. Rather than a per-provider transcode, text_to_speech_tool now sniffs the audio magic bytes once after every synthesis and repairs any MP3/WAV-bytes-in-.ogg centrally (in-place ffmpeg transcode to real Ogg/Opus, honest-extension rename fallback when ffmpeg is missing) — covering this provider and all others, including command providers and plugins. Your diagnosis of the container mismatch was correct and helped shape the central fix — thanks for the contribution.