fix(tts): transcode xAI TTS output to Opus when .ogg requested (#57213)

[AlexFucuson9]

Problem

xAI /v1/tts API only supports mp3/wav codecs. When caller requests .ogg output (Telegram auto voice replies), MP3 bytes silently written into .ogg file -> broken voice bubbles.

This is xAI instance of known bug class (#57048, #54589). Existing fixes (#55278, #57069, #56873) target other providers but none covers xAI.

Root Cause

tools/tts_tool.py::_generate_xai_tts:
codec = wav if .wav else mp3 -- no .ogg handling. xAI API has no native Opus codec.

Fix

When output_path ends with .ogg:

1. Write API response to temp .mp3 file
2. Transcode to Opus/OGG via existing _convert_to_opus (ffmpeg)
3. Clean up temp file, return .ogg path
4. Fall back to mp3 if ffmpeg unavailable

Same pattern as dispatcher-level opus conversion for other providers.

Closes #57213

[alt-glitch]

This was generated by AI during triage.
Duplicate of #57222 — same fix at the same site: both edit tools/tts_tool.py::_generate_xai_tts to synthesize to a temp MP3 then transcode to Opus/OGG via _convert_to_opus for .ogg targets. #57222 (2026-07-02T16:29) is the earliest open xAI-specific fix; #57250 was already deduped into it. Consolidating on #57222.

[teknium1]

Thanks for targeting a real xAI/Telegram incompatibility: current main still writes xAI's MP3 response into a requested .ogg path at tools/tts_tool.py:1370-1413, while Telegram auto replies request .ogg at gateway/run.py:13300-13310.

Problems

• tools/tts_tool.py:1300 returns the OGG beside the temporary MP3. _convert_to_opus() derives that name from its input (tools/tts_tool.py:920), so it is not the requested output path.
• The xAI dispatch at tools/tts_tool.py:2443-2445 ignores that return value, then validates the original .ogg path at tools/tts_tool.py:2528-2533. The proposed successful-conversion path therefore reports no output.
• This PR changes only tools/tts_tool.py; please add a regression test for this .ogg path.

Suggested changes

• Move/replace the converted OGG onto output_path before returning, or propagate the actual returned path through the dispatcher and correctly handle the MP3 fallback.
• Add a test that verifies the requested .ogg path is the converted output, including ffmpeg-unavailable behavior.

This is an automated hermes-sweeper review.

[teknium1]

_convert_to_opus() derives an OGG beside this temporary MP3, but text_to_speech_tool() ignores the xAI generator's return value (tools/tts_tool.py:2443-2445) and still validates the original requested .ogg path. Move the converted file to output_path before returning, or propagate the returned path through the dispatcher.

[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.