fix(local-asr): 切走/删模型/手动释放时真正归还 RAM

[appergb]

User description

Summary

修复"按了释放按钮内存没掉"以及一并查出的两个相关泄漏点。三处都是同一主题：本地 Qwen3-ASR engine 该走时却没真走。

1. macOS 主动归还内核（用户报告主因）

按"释放"按钮后 Arc<QwenAsrEngine> 已经 drop、qwen_free 也跑完了，但活动监视器看 RSS 几乎不动。原因：encoder weights 是 f32 malloc 出来的几百 MB，free() 后 macOS libmalloc 把页留在 freelist 上不归还内核。decoder bf16 走 mmap，munmap 已立即生效，所以会"掉一部分但不掉大头"。

修：在 LocalAsrCache::release_now / release_if_idle 真 drop 完 engine 后调 malloc_zone_pressure_relief(NULL, 0)，让 libmalloc 主动归还。

2. 切回云端时本地引擎不释放

set_active_asr_provider 当前只在切到 local 时预加载，切回云端时没有镜像处理。结果：用户切走后 1.2GB+ RAM 一直留到进程退出，下次再走 dictation 是云端路径，根本不会触发 schedule_local_asr_release。

修：切到非 local-qwen3 时 coord.release_local_asr_engine()。

3. 删模型时不释放内存里同 id 引擎

local_asr_delete_model 直接 delete_model(id) 删盘，不看内存里有没有加载这个 id。删完后 mmap 还指着已 unlink 的文件，RAM 还在。

修：local_asr_delete_model 注入 coord，删之前若 loaded_model_id == 要删的 id 先 release。

平台范围

Windows 端本地 ASR 本期不上线（issue #256），改动全部在 #[cfg(target_os = "macos")] 块内或只影响 macOS 唯一持有引擎的代码路径。

Test plan

• 设置页 → 加载本地模型 → 按"释放" → 日志出现 release engine ... on demand + malloc_zone_pressure_relief freed ~N bytes，活动监视器 RSS 立即下降几百 MB（含 encoder malloc 部分）
• 设置页 → 加载本地模型 → 切到火山引擎 → 上述两条日志同样出现，RAM 立即释放
• 设置页 → 加载本地模型 → 删除该模型 → 释放日志出现，磁盘文件被删除
• 跑一次正常听写，确认 dictation 路径未受影响（cache hit / schedule_local_asr_release 行为不变）
• 跑 cargo check --manifest-path src-tauri/Cargo.toml 通过

---

PR Type

Bug fix

---

Description

• Manually release engine now reclaims memory on macOS

  • Added malloc_zone_pressure_relief after engine drop to return pages to kernel
  • Heavy drop and pressure relief moved outside the lock for better concurrency

• Switching ASR provider frees local engine immediately

  • set_active_asr_provider now calls release_local_asr_engine when selecting cloud providers

• Deleting model safely releases engine before file removal

  • local_asr_delete_model receives coordinator, releases engine if loaded model matches the one being deleted

---

Diagram Walkthrough

flowchart LR
  A["Manual Release"] -- "drop + pressure_relief" --> B["RSS Drops"]
  C["Switch to Cloud"] -- "release_local_asr_engine" --> D["Free 1.2GB+"]
  E["Delete Model"] -- "release if loaded" --> F["Safe unlink"]

Loading

File Walkthrough

Relevant files

Bug fix

cache.rs macOS memory pressure relief after ASR engine drop              openless-all/app/src-tauri/src/asr/local/cache.rs Added pressure_relief_macos() to force macOS malloc to return freed pages to the kernel Moved engine drop and pressure relief after lock release to improve concurrency +37/-12  commands.rs Release engine on ASR provider switch and model deletion  openless-all/app/src-tauri/src/commands.rs set_active_asr_provider: release local engine when switching to a non-local provider local_asr_delete_model: accept coordinator, release engine if loaded model is targeted for deletion +15/-2

---

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis ❌ 256 - Not compliant Non-compliant requirements: Windows local ASR option is not grayed out or tooltiped in this PR. No UI changes for keeping Windows users on cloud-only ASR paths are present here. The PR does not address the Windows-only scope of the ticket.

⏱️ Estimated effort to review: 2 🔵🔵⚪⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ No major issues detected