perf(macos): 保留 MLX 原生构建缓存

.github/actions/cache-macos-mlx/action.yml

@@ -0,0 +1,42 @@
+name: Cache macOS MLX native build
+description: Preserve vendored MLX CMake output before rust-cache removes in-tree path dependencies.
+inputs:
+  profile:
+    description: Cargo output profile directory (debug or release)
+    required: true
+runs:
+  using: composite
+  steps:
+    - name: Fingerprint MLX native toolchain
+      id: native
+      shell: bash
+      working-directory: openless-all/app
+      env:
+        OPENLESS_MLX_PROFILE: ${{ inputs.profile }}
+      run: |
+        set -euo pipefail
+        case "$OPENLESS_MLX_PROFILE" in
+          debug|release) ;;
+          *) echo "::error::Unsupported MLX cache profile"; exit 1 ;;
+        esac
+        fingerprint=$({
+          rustc -vV
+          xcrun clang --version
+          xcrun --sdk macosx --show-sdk-version
+          cmake --version
+          git -C src-tauri/vendor/qwen3-asr-rs rev-parse HEAD
+          git -C src-tauri/vendor/qwen3-asr-rs/mlx-c rev-parse HEAD
+          for name in CARGO_PROFILE_DEV_DEBUG CARGO_PROFILE_TEST_DEBUG CARGO_PROFILE_RELEASE_CODEGEN_UNITS CARGO_PROFILE_RELEASE_STRIP RUSTFLAGS CARGO_ENCODED_RUSTFLAGS CC CXX CFLAGS CXXFLAGS SDKROOT DEVELOPER_DIR MACOSX_DEPLOYMENT_TARGET CMAKE_GENERATOR CMAKE_TOOLCHAIN_FILE; do
+            printf '%s=%s\n' "$name" "${!name-}"
+          done
+        } | shasum -a 256 | awk '{print $1}')
+        echo "fingerprint=$fingerprint" >> "$GITHUB_OUTPUT"
+
+    # Call this action AFTER rust-cache: post actions run in reverse order,
+    # so MLX is saved before rust-cache prunes src-tauri/vendor path packages.
+    # Cache only CMake output, not Cargo fingerprints: Cargo still reruns the
+    # build script and CMake validates its native inputs on every clean checkout.
+    - uses: actions/cache@v4
+      with:
+        path: openless-all/app/src-tauri/target/${{ inputs.profile }}/build/qwen3-asr-rs-*/out
+        key: macos-mlx-v1-${{ runner.arch }}-${{ inputs.profile }}-${{ steps.native.outputs.fingerprint }}-${{ hashFiles('openless-all/app/src-tauri/Cargo.toml', 'openless-all/app/src-tauri/Cargo.lock', 'openless-all/app/src-tauri/tauri*.json', '.cargo/config.toml', '.github/actions/cache-macos-mlx/action.yml') }}

.github/workflows/ci.yml

@@ -252,6 +252,12 @@ jobs:
             openless-all/app -> target
             openless-all/app/src-tauri -> target
 
+      - name: Cache macOS MLX native build
+        if: runner.os == 'macOS' && runner.arch == 'ARM64'
+        uses: ./.github/actions/cache-macos-mlx
+        with:
+          profile: debug
+
       - name: Prepare Windows Sherpa static libraries
         if: runner.os == 'Windows'
         shell: pwsh
@@ -388,6 +394,11 @@ jobs:
           workspaces: |
             openless-all/app/src-tauri -> target
             openless-all/app/src-tauri/backend-tests -> target
+      - name: Cache macOS MLX native build
+        if: runner.arch == 'ARM64'
+        uses: ./.github/actions/cache-macos-mlx
+        with:
+          profile: debug
       - run: npm ci
       - run: npm run build
       - name: Check Tauri backend with Rust 1.88 MSRV

.github/workflows/release-tauri.yml

@@ -92,6 +92,12 @@ jobs:
           key: macos-release-v1
           workspaces: 'openless-all/app/src-tauri -> target'
 
+      - name: Cache macOS MLX native build
+        if: matrix.updater-arch == 'aarch64'
+        uses: ./.github/actions/cache-macos-mlx
+        with:
+          profile: release
+
       - name: Prepare Windows Sherpa static libraries
         if: matrix.platform == 'windows-latest'
         working-directory: 'openless-all/app'

docs/macos-build-performance.md

@@ -31,6 +31,8 @@ ARM/Intel 的整个打包步骤分别约 15 分 17 秒、15 分 08 秒。主要
 
 macOS 检查统一关闭 dev/test debug info，分别缓存 Core、Host、backend-tests 的实际 target 目录。stable、MSRV、release 缓存分开，避免不同工具链和 profile 的产物互相挤占。
 
+MLX 的 CMake 输出另用 [cache-macos-mlx](../.github/actions/cache-macos-mlx/action.yml) 保存。实际使用的 `rust-cache` [源码](https://github.com/Swatinem/rust-cache/blob/6323deb102c322ba6fcbdcafc7e3dddab59af2b6/src/workspace.ts) 排除 workspace 目录内的 path 依赖，导致 `src-tauri/vendor/qwen3-asr-rs` 的原生输出在 post 阶段被清理；首轮验证中，命中 Cargo 缓存仍重建 MLX 约 7 分 17 秒。独立缓存步骤放在 `rust-cache` 后，利用 post 的逆序执行先保存原生输出。缓存按架构、profile、Rust/Clang/CMake/SDK、子模块提交、编译环境和 manifest/lock/config 隔离，无跨 key 的模糊回退。只保存 CMake `out`，不保存 Cargo freshness 指纹，下一轮仍执行 build script 与 CMake 输入校验。
+
 `scripts/macos-build-env.sh` 为 macOS 打包默认设置 `CARGO_PROFILE_RELEASE_CODEGEN_UNITS=16`，保留 `opt-level=3`、thin LTO 和 unwind；环境变量可以显式覆盖为其他值。参数取舍依据 [Cargo profiles](https://doc.rust-lang.org/cargo/reference/profiles.html#codegen-units)：更多 codegen units 允许更快的并行代码生成，可能影响最终体积或优化效果。共享 `Cargo.toml` 不修改。CI 在恢复缓存前加载同一环境，本地 `build-mac.sh` 也加载它。
 
 MLX 清理只比较含非空 metallib 的输出目录，保留 build-script 可执行文件。构建前删除本次架构的旧 app、DMG 和 updater，保留 Cargo 编译缓存；Tauri 非零退出直接失败。因此热构建复用旧时间戳二进制时仍能正确打包，失败时也不会接受旧安装包。现有用途声明、签名、公证和 MLX 包内容校验继续执行。