[codex] Preserve explicit model capabilities in settings

[LeonLiuMercaso]

Summary

• Preserve explicit model capabilities selected in Settings instead of letting inferred model profiles override them.
• Default likely ASR candidates, such as qwen3-asr-*, to the transcription capability when selected from fetched model candidates.

Root Cause

When a user manually added an ASR model and selected “转录模型”, toAiModelDescriptor still preferred forced capabilities inferred from broad model profile rules. A generic qwen3 profile could therefore rewrite qwen3-asr-flash-* back to chat, causing the transcription model dropdown to stay empty after saving.

User Impact

Users configuring Alibaba Bailian / DashScope ASR models could not persist the model as a transcription model, even after choosing the transcription capability in the UI.

Validation

• npx tsc --noEmit
• git diff --check

[LeonLiuMercaso]

已按要求改用 leonliu76-web 账号重新提交中文 PR：#37。这个英文 PR 关闭以避免重复。