[iOS] Add preset X-ASR flow to Qwen3.5 chat

[ring-zl]

Summary

• add a preset audio chip beside the existing image presets in the Qwen3.5-2B demo
• run the bundled sample through X-ASR (Zipformer2, MNN weight INT8 block64) on CPU before submitting the normalized transcript to the LLM
• show the playable waveform first, then the transcript and ASR-only latency/RTF in compact text
• keep model initialization time out of the user-visible metric while retaining it in diagnostics
• add generation-based result invalidation, a permanent native worker thread, asset SHA-256 validation, and a silent signed-device probe

Asset delivery

Model weights, the sample recording, and the generated arm64 libsherpa-mnn.a are intentionally excluded from Git. The internal test package is:

mnn-ios-xasr-int8-test-assets-v1-final.zip

SHA-256:

a8eb55870620a5195b458aa9c316e1cbee0a1af600002239095eade64aeffdcf

Please confirm redistribution rights for X-ASR and the sample recording before publishing the archive. apps/iOS/MNNLLMChat/PRESET_ASR.md documents installation and exact per-file hashes.

Verification

• clean signed-device Debug build: PASS
• device: iPhone 13 Pro (iPhone14,2)
• silent production-path probe: PASS; no microphone capture or speaker playback
• sample duration: 2.136875 s
• normalized transcript: 给我介绍一下杭州市的美食
• X-ASR inference: approximately 0.294 s
• RTF: approximately 0.138 (approximately 7.3x real time)
• interactive audio/transcript/LLM flow: manually accepted on device
• package extract + six-file SHA-256 verification: PASS

Not tested: corpus-level CER/WER, repeated-turn/cancellation stress matrix, memory peak/Jetsam headroom, simulator or iPad.

Notes

The first commit removes the unavailable cmark-gfm-bin target from the vendored SwiftPM manifest. The branch does not contain its declared bin directory, so a clean dependency resolution otherwise fails before the app target is built.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

周林（482152） seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[ring-zl]

@v0jiuqi 你好，这个 PR 基于你提交的 feature/ios-app-dev，补充了预设音频 → X-ASR → Qwen3.5-2B 的真机流程。

我已经在 iPhone 13 Pro 上完成无声 production-path probe 和交互测试。测试音频 2.136875 秒，X-ASR 推理约 0.294 秒，RTF 约 0.138，识别结果为“给我介绍一下杭州市的美食”。模型使用 X-ASR Zipformer2 的 MNN weight INT8 block64 版本。

模型、样例音频和 arm64 libsherpa-mnn.a 已整理成独立内部测试包，没有提交到 GitHub。包 SHA-256：

a8eb55870620a5195b458aa9c316e1cbee0a1af600002239095eade64aeffdcf

请告知适合的内部交付渠道，或确认应该上传到哪个 alinnprivate 模型位置；确认再分发权限前我不会公开上传该包。部署路径、逐文件 hash 和 silent probe 方法已写在 apps/iOS/MNNLLMChat/PRESET_ASR.md。