fix(speech): route flash ASR to sync endpoint

[LauraGPT]

Fixes #146.

What changed

• route fun-asr-flash and qwen-audio-3.0-asr-flash model families to the synchronous multimodal generation endpoint
• keep fun-asr and *-filetrans models on the existing asynchronous transcription endpoint
• build the documented input_audio request body, infer the audio format, and preserve language/vocabulary options
• reject async-only flags and multiple inputs early for Flash models
• document model selection and add endpoint/request/response regression coverage

Root cause

speech recognize previously sent every model to /api/v1/services/audio/asr/transcription. Alibaba Cloud's HTTP API requires the Flash families to use /api/v1/services/aigc/multimodal-generation/generation synchronously, which caused the reported HTTP 400 URL error.

Validation

• pnpm run sync:skill-assets
• pnpm exec vp check
• pnpm exec vp test --run
• git diff --check
• commands package: 325 passed, 174 environment-gated skipped
• core and commands builds passed
• local HTTP contract test verifies endpoint, headers, request body, stdout, and --out response persistence

A live Bailian request was not run because no DashScope API key is configured on the test host; the request contract is based on the official HTTP API documentation and exercised against a local HTTP server.

[XXPermanentXX]

Thanks for the contribution! PR #149 addresses the same issue with broader ASR model routing and test coverage, including the Flash and Filetrans model families. We’ve incorporated the relevant endpoint and request-handling improvements there, so we plan to merge #149 and close this PR to avoid duplicate changes. Really appreciate your investigation and contribution!