fix: strip Qwen3-ASR output prefix from transcriptions

[malaiwah]

Qwen3-ASR (via vLLM/sglang OpenAI-compatible endpoint) returns transcriptions with a structured prefix:

language English<asr_text>Hello. This is the actual text.

This prefix was passed verbatim to the LLM as the user's voice message. On short utterances it could confuse the model.

Fix: strip everything up to and including <asr_text> in _extract_transcript_text(). The marker is XML-like and cannot appear in natural speech, so false positives are impossible. The if '<asr_text>' in text guard makes this a no-op for non-Qwen providers (Whisper, Groq, etc.).

[teknium1]

Thanks for identifying the Qwen3-ASR response shape. The premise remains valid on current main: tools/transcription_tools.py:1787 returns string responses unchanged except for whitespace, and the OpenAI-compatible path calls this helper at tools/transcription_tools.py:1360.

Problems

• The diff's right-side tools/transcription_tools.py:105 replaces provider-aware model resolution with a raw flat stt.model lookup. This is unrelated to prefix stripping and drops documented behavior from the PR base; it should not be part of the salvage.
• The diff has no regression test. Current OpenAI transcription coverage in tests/tools/test_managed_media_gateways.py:305-346 covers ordinary text/JSON responses but not the Qwen marker.

Suggested changes

• Apply only the marker normalization to current main's _extract_transcript_text() (tools/transcription_tools.py:1784) and preserve the current configuration architecture.
• Add direct regression cases for string, .text, and dictionary text responses containing <asr_text>.

Automated hermes-sweeper review.

[teknium1]

This unrelated replacement drops the provider-aware model lookup. Please omit this hunk when salvaging the Qwen transcript normalization; the prefix fix does not require changing model-resolution behavior.

[LauraGPT]

I refreshed this valid fix on current main in #65534.

The replacement keeps the current provider-aware STT configuration behavior, applies <asr_text> normalization to string / SDK .text / dictionary responses, and adds direct regressions for all three shapes. The original contribution remains credited in the commit trailer.

Fresh validation on the replacement head: 235 related tests passed, plus Ruff, changed-range formatting, Python compilation, and diff checks. Since this branch is now reported as conflicting, centralizing the current-main review in #65534 avoids carrying the stale unrelated config diff forward.