fix: drop stale language-hint caveats in Qwen3 ASR

docs/models/qwen3-asr-0.6b.md

@@ -7,7 +7,7 @@ Offline multilingual speech-to-text. An 18-layer bidirectional audio encoder
 feeds a 28-layer Qwen3 causal LM with audio-token injection (fused
 audio+text sequence, no cross-attention). Auto-detects the audio's language
 across 30 languages and emits the transcript in that language. Takes a
-16 kHz mono WAV; explicit language hints are not supported at this time.
+16 kHz mono WAV and produces a transcript.
 <!-- /catalog -->
 
 ## What it's for
@@ -113,16 +113,6 @@ If your audio is not already 16 kHz mono WAV, convert it first:
 ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav
 ```
 
-## Public API caveat — language hints
-
-This port accepts `params.language == NULL` (auto-detect) and
-**rejects any explicit language hint** with
-`TRANSCRIBE_ERR_UNSUPPORTED_LANGUAGE`. The `capabilities.languages`
-list documents the 30 languages the model can auto-detect, not a set
-of caller-settable hints. Rendering caller-supplied hints into the
-chat template is tracked as follow-up work; see the family note at
-`docs/porting/families/qwen3_asr.md` for details.
-
 ## Performance
 
 ### Apple M4 Max

docs/models/qwen3-asr-1.7b.md

@@ -8,8 +8,7 @@ Offline multilingual speech-to-text. Same audio-LLM architecture as the
 audio-token injection), wider: encoder `d_model=1024` (16 heads), LM
 `hidden_size=2048`, `intermediate_size=6144`. Auto-detects the audio's
 language across 30 languages and emits the transcript in that language.
-Takes a 16 kHz mono WAV; explicit language hints are not supported at
-this time.
+Takes a 16 kHz mono WAV and produces a transcript.
 <!-- /catalog -->
 
 ## What it's for
@@ -107,13 +106,6 @@ If your audio is not already 16 kHz mono WAV, convert it first:
 ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav
 ```
 
-## Public API caveat — language hints
-
-Same contract as the 0.6B: `params.language == NULL` runs auto-detect;
-any explicit hint returns `TRANSCRIBE_ERR_UNSUPPORTED_LANGUAGE`. See the
-0.6B doc and the family note at `docs/porting/families/qwen3_asr.md` for
-the rationale and the planned follow-up.
-
 ## Performance
 
 ### Apple M4 Max

docs/porting/families/qwen3_asr.md

@@ -132,14 +132,6 @@ Bridge validation:
 Things the first port intentionally does not do; tracked as follow-
 up work rather than shipped-and-broken.
 
-- **Language hinting is rejected.** `transcribe_run_params.language == NULL`
-  is the supported mode and triggers the model's built-in auto-detect
-  (it prefixes the transcript with `language X`, which we strip before
-  returning). Any non-null hint, including a language in the
-  capability list, returns `TRANSCRIBE_ERR_UNSUPPORTED_LANGUAGE`. The
-  `caps.languages` list documents the model's auto-detect coverage,
-  not what callers may hint — rendering caller-supplied hints into the
-  chat template is a future change.
 - **Streaming** (`stream_transcribe` / chunk rollback) is out of scope
   for this port. Upstream Qwen3-ASR may be architecturally usable in a
   streaming mode, but the transcribe.cpp library does not expose or

scripts/hf_cards/qwen3-asr-0.6b.yaml

@@ -25,7 +25,7 @@ summary: |
   feeds a 28-layer Qwen3 causal LM with audio-token injection (fused
   audio+text sequence, no cross-attention). Auto-detects the audio's language
   across 30 languages and emits the transcript in that language. Takes a
-  16 kHz mono WAV; explicit language hints are not supported at this time.
+  16 kHz mono WAV and produces a transcript.
 
 wer:
   notes: |

scripts/hf_cards/qwen3-asr-1.7b.yaml

@@ -26,8 +26,7 @@ summary: |
   audio-token injection), wider: encoder `d_model=1024` (16 heads), LM
   `hidden_size=2048`, `intermediate_size=6144`. Auto-detects the audio's
   language across 30 languages and emits the transcript in that language.
-  Takes a 16 kHz mono WAV; explicit language hints are not supported at
-  this time.
+  Takes a 16 kHz mono WAV and produces a transcript.
 
 wer:
   notes: |

tests/fixtures/make_gguf_fixtures.py

@@ -1757,11 +1757,7 @@ def emit_fixtures(out_dir: Path) -> None:
                 _pack_kv_string("stt.variant",          "qwen3-asr-toy"),
                 _pack_kv_string("tokenizer.chat_template",
                                 QWEN3_ASR_CHAT_TEMPLATE),
-                # Capabilities: match the family note. The loader reads
-                # general.languages as factual detection coverage; the
-                # run() handler rejects any explicit language hint with
-                # TRANSCRIBE_ERR_UNSUPPORTED_LANGUAGE until the prompt
-                # renderer honors them (Phase 1.1 Option A).
+                # Capabilities: match the family note.
                 _pack_kv_bool("stt.capability.lang_detect", True),
                 _pack_kv_array_string(
                     "general.languages",