fix(stt): improve Soniqo Qwen transcript coverage

[ComputelessComputer]

Normalize Soniqo language hints to ISO codes and pad short Qwen audio chunks so VAD-split speech is not dropped.

---

Note

Medium Risk
Behavior changes in Soniqo transcription (language hint mapping and Qwen3 chunk handling) may alter transcript output and edge-case coverage, but does not touch security-sensitive logic.

Overview
Improves Soniqo batch transcription by normalizing language hints from BCP-47 (e.g. en-US) to an ISO-639 code (e.g. en) via soniqo_language_hint, with a regression test to lock the behavior.

Updates the Soniqo Swift bridge’s Qwen3 path to pad sub-1-second audio up to a minimum sample count and run chunking/transcription on the padded buffer, preventing very short segments from returning empty transcripts.

^{Reviewed by Cursor Bugbot for commit 8058747. Bugbot is set up for automated code reviews on this repo. Configure here.}

[netlify]

✅ Deploy Preview for old-char canceled.

Name	Link
🔨 Latest commit	8058747
🔍 Latest deploy log	https://app.netlify.com/projects/old-char/deploys/6a0c8a7895e89b0008cb5ec0

[ComputelessComputer]

Closing as superseded by #5284. Since Soniqo Omnilingual and Qwen are being removed from the selectable local STT model list, we do not need to carry improvements specific to Qwen transcription right now.