perf(ci): cache cargo deps and drop the duplicate Windows full test run (#404)

.github/workflows/ci.yml

@@ -38,6 +38,12 @@ jobs:
           rustup toolchain install
           rustup show active-toolchain
 
+      # #404: dependency build artifacts + cargo registry. Must follow Setup
+      # Rust: the cache key includes the active rustc. Only dependencies are
+      # cached; the workspace crate is always rebuilt.
+      - name: Cache cargo registry and dependency builds
+        uses: Swatinem/rust-cache@6323deb102c322ba6fcbdcafc7e3dddab59af2b6 # v2.9.2
+
       - name: Install ripgrep
         shell: pwsh
         run: choco install ripgrep --version 14.1.0 -y --no-progress
@@ -384,6 +390,12 @@ jobs:
           rustup toolchain install
           rustup show active-toolchain
 
+      # #404: see the matching step in windows-build-test-package. The cache
+      # key includes the job id and runner OS/arch, so each matrix leg keeps
+      # its own entry.
+      - name: Cache cargo registry and dependency builds
+        uses: Swatinem/rust-cache@6323deb102c322ba6fcbdcafc7e3dddab59af2b6 # v2.9.2
+
       - name: Install ripgrep
         shell: pwsh
         run: |
@@ -440,7 +452,13 @@ jobs:
         shell: pwsh
         run: cargo check --locked
 
+      # #404: on Windows this exact command already runs in
+      # windows-build-test-package (same runner image, same toolchain, same
+      # `-p code-intel --locked` test set), so the matrix leg skips it instead
+      # of executing the full suite a second time per trigger. The packaged
+      # install_smoke tests below still compile and run here on every OS.
       - name: Rust tests and core cross-contract gate
+        if: runner.os != 'Windows'
         shell: pwsh
         run: cargo test -p code-intel --locked
 

.github/workflows/pr-gate.yml

@@ -18,11 +18,10 @@ name: code-intel-pr-gate
 # inflated percentile as a defect (see the sample-exclusion comment there);
 # a moving denominator is the same defect from another source.
 #
-# Toolchain/cache/action choices are mirrored from ci.yml on purpose: ci.yml
-# has no dependency-cache action (just `rustup toolchain install` reading
-# rust-toolchain.toml on an uncached runner), so this workflow does not add
-# one either — introducing caching here would be a new third-party action
-# ci.yml itself does not use.
+# Toolchain/action choices are mirrored from ci.yml on purpose (`rustup
+# toolchain install` reading rust-toolchain.toml). ci.yml adopted the pinned
+# Swatinem/rust-cache in #404; this workflow does not use it yet, so its jobs
+# still build on an uncached runner.
 
 on:
   pull_request:

CHANGELOG.md

@@ -42,6 +42,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - 独立审计确认 4 个 `templates/` 文件（`understanding-report.md`、`minimax-deploy-checklist.md`、`dependency-audit.md`、`idea-file.md`）在 .md/.json/.yml/.toml/.rs/.py 全仓与 orchestration digest pin 中零引用，予以删除；`templates/sentrux-rules.example.toml` 因被 README/CHANGELOG 实际引用而保留（#357/#358）。
 - **session-observability 原型经真实试跑后判定不达晋升门槛，移除**（#359/#360）：安装 cosmtrek/mindwalk（钉版本），用真实 Claude Code session 生成的 trace 对照该 session 提交前后状态生成的规范 `sentrux-hotspots.json` 实测联接——5 条晋升标准中"trace 目标能联接上 Sentrux hotspot 记录"实测 0 联接（session 只碰了 2 个小测试文件，均不在旧 top-30 复杂度截断内），未达 IDEA.md 定的晋升门槛（"仅当真实 trace 产出有用的路径联接且 timeline 暴露 review 相关信号才晋升"），按负结果移除原型，issue 记录结论与成因供后续复议。
 - **`sentrux_gate` 的 `quality_signal` 从 `10000 - penalty` 代理分数改为上游兼容的 Quality Signal，并输出五个根因（#385，父级 #285）**：新增 `sentrux_quality_signal.rs` 计算内核，按上游文档（https://sentrux.dev/docs/quality-signal/）与固定源码版本（`6f8ff3c14b0423e4b58f42d1813d4d5f7fdc1d11`）移植 modularity（Newman Q）、acyclicity、depth、equality（Gini）、redundancy 五因子归一化公式与几何平均聚合，显式记录并保留固定源码相对文档页多出的 `max(0.01)` 逐因子下限（`FORMULA_VERSION`）。原生扫描器此前无跨语言解析的文件依赖图，`measure_project` 新增有界启发式导入目标解析（引号相对路径 + 唯一文件名回退）喂给 modularity/acyclicity/depth；equality 采用上游自身定义的"无函数级 CC 数据时退回文件行数"回退路径（本仓从未追踪过函数级复杂度）；redundancy 目前只实现 duplicate 一半（跨文件规范化内容比对），dead-function 检测未实现——本仓 `cargo check` 的 ~100 条 dead-code 警告本就是 `#[path]` 重复编译的编译器产物而非真实死代码信号，朴素的"从未被调用"启发式对本仓会产生大量假阳性，因此诚实标记为 `completeness: "partial"` 并在 `note` 字段说明，不伪造 `dead_count = 0`。`scan`/`health` 新增 `quality_signal_detail`（`schema`/`formula_version`/`provider_version`/`bottleneck`/`completeness`/`root_causes.<metric>.{score,raw,completeness}`）。旧代理分数与新分数语义不兼容，`BASELINE_SCHEMA` 随之升到 v6（`ENGINE_VERSION` 3.0.0），旧 v5 基线按既有 `baseline_engine_mismatch` 机制拒绝直接比较并给出机器可读原因，避免伪造 before/after；本仓自身 `.sentrux/baseline.json` 已用新引擎重新保存（非回归抑制，是既有"改变多 session 行为口径必须重新定基线"惯例的直接后果）。见 DR-0011。
+- **CI：cargo 依赖缓存 + 去掉 Windows 重复全量测试（#404）**：`ci.yml` 两个编译 job（`windows-build-test-package`、`cross-platform-smoke` 矩阵）在 `Setup Rust` 之后接入按 commit SHA 钉住的 `Swatinem/rust-cache`（v2.9.2，只缓存 cargo registry 与依赖构建产物，key 含 job id、runner OS/arch、rustc 与 lockfile）。`cross-platform-smoke (windows-latest)` 不再执行 `cargo test -p code-intel --locked`——同一条命令在同一镜像上已由 `windows-build-test-package` 执行；该矩阵腿独有的打包安装拓扑检查（bootstrap.py 安装、`install_smoke` 两条 `--ignored` 测试、stale-home、Sentrux closure、PATH 持久化断言，DR-0001）全部保留。必需检查名不变。
 
 ## [0.7.2-beta.6] — 2026-08-19