ci(judge-e2e): skip the engine on JS-only changes and cut its build from ~37 min

.github/workflows/judge-e2e.yml

@@ -53,29 +53,34 @@ jobs:
         with:
           persist-credentials: false
           ref: ${{ github.sha }}
+          # The PR merge commit plus its first parent (the base tip) for the diff below.
+          fetch-depth: 2
+
+      # packages/console-ui (types only) and the root pnpm files feed the UI
+      # checks but nothing the Rust engine suite exercises, so a PR touching
+      # only those skips the Rust engine half. Keep this regex in sync with
+      # on.pull_request.paths minus those JS-only entries.
+      - name: Decide whether the engine suite runs
+        id: scope
+        run: |
+          changed="$(git diff --name-only HEAD^1 HEAD)"
+          if grep -qE '^(judge(-laya|-typesafe|-semif|-decider|-clef|-openai)?/|crates/(llama-runtime|llama-native|judge-contract|judge-provider|config-client|console-ui|worker-paths)/|rust-toolchain\.toml$|\.cargo/|\.github/scripts/judge-|\.github/workflows/judge-e2e\.yml$)' <<<"$changed"; then
+            echo "engine=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "::notice::Only JS-side shared files changed; skipping the Rust engine suite."
+          fi
 
       - name: Prepare reports
         run: |
           mkdir -p "$RUNNER_TEMP/judge-e2e-reports"
           echo "JUDGE_E2E_REPORT_DIR=$RUNNER_TEMP/judge-e2e-reports" >> "$GITHUB_ENV"
 
       - name: Install Rust toolchain
+        if: steps.scope.outputs.engine == 'true'
         uses: dtolnay/rust-toolchain@6c977a6ca4077a0ceb28ffbe03f59d46e9ac8772 # v1
         with:
           toolchain: 1.97.1
 
-      - name: Cache cargo registry and builds
-        uses: Swatinem/rust-cache@6323deb102c322ba6fcbdcafc7e3dddab59af2b6 # v2
-        with:
-          workspaces: |
-            judge
-            judge-laya
-            judge-typesafe
-            judge-clef
-            judge-openai
-            crates/judge-contract
-            crates/judge-provider
-
       - name: Setup pnpm
         # Version comes from the root package.json packageManager field.
         uses: pnpm/action-setup@fc06bc1257f339d1d5d8b3a19a8cae5388b55320 # v5
@@ -112,13 +117,8 @@ jobs:
           # Reuse these assets in both Cargo feature variants.
           echo 'SKIP_UI_BUILD=1' >> "$GITHUB_ENV"
 
-      # judge-semif, judge-decider, judge-laya and judge-clef compile llama.cpp
-      # through crates/llama-native (curl, tar, patch, cmake; no libclang) and
-      # the runtime-loaded Vulkan backend module (headers + glslc).
-      - name: Provision llama.cpp build tools
-        run: sudo apt-get update && sudo apt-get install -y cmake libvulkan-dev glslc spirv-headers
-
       - name: Install pinned iii engine
+        if: steps.scope.outputs.engine == 'true'
         env:
           GITHUB_TOKEN: ${{ github.token }}
           BIN_DIR: ${{ runner.temp }}/judge-engine/bin
@@ -131,7 +131,18 @@ jobs:
           test "$(awk '{print $NF}' "$JUDGE_E2E_REPORT_DIR/engine-version.log")" = "${III_RELEASE_TAG#iii/v}"
           echo "III_ENGINE_BIN=$BIN_DIR/iii" >> "$GITHUB_ENV"
 
+      # Compile time was the whole cost: four from-scratch llama.cpp builds with
+      # every CPU variant and Vulkan, plus eight workspaces each building the
+      # same deps. One target dir builds each shared unit once across the
+      # suites (judge-laya's opt-level 3 still gets its own llama.cpp), and the
+      # tiny models run on one CPU module. The hosted image ships cmake, curl,
+      # tar and patch. No rust-cache: PR-scoped caches never reach another PR,
+      # and it purges path crates (llama-native's llama.cpp) anyway.
       - name: Real-engine tests (11 local mocked cases)
+        if: steps.scope.outputs.engine == 'true'
+        env:
+          CARGO_TARGET_DIR: ${{ runner.temp }}/judge-e2e-target
+          III_LLAMA_CPP_MINIMAL: '1'
         run: bash .github/scripts/judge-e2e.sh
 
       - name: Upload test logs

crates/llama-native/build.rs

@@ -101,6 +101,7 @@ fn sha256_hex(path: &Path) -> Option<String> {
 fn llama() {
     println!("cargo:rerun-if-changed=native");
     println!("cargo:rerun-if-env-changed=III_LLAMA_CPP_TARBALL");
+    println!("cargo:rerun-if-env-changed=III_LLAMA_CPP_MINIMAL");
     let out = PathBuf::from(env::var("OUT_DIR").expect("cargo sets OUT_DIR"));
     let src = llama_source(&out);
     let os = env::var("CARGO_CFG_TARGET_OS").expect("cargo sets the target OS");
@@ -126,11 +127,18 @@ fn llama() {
         .define("GGML_OPENMP", "OFF")
         .define("BUILD_SHARED_LIBS", if dynamic { "ON" } else { "OFF" });
     if dynamic {
+        // III_LLAMA_CPP_MINIMAL=1 (test-only builds such as Judge E2E): one CPU
+        // module and no Vulkan, a fraction of the build time. Never ship it.
+        let full = if env::var("III_LLAMA_CPP_MINIMAL").as_deref() == Ok("1") {
+            "OFF"
+        } else {
+            "ON"
+        };
         fs::create_dir_all(&backends).expect("creating the backends directory");
         cmake
             .define("GGML_BACKEND_DL", "ON")
-            .define("GGML_CPU_ALL_VARIANTS", "ON")
-            .define("GGML_VULKAN", "ON")
+            .define("GGML_CPU_ALL_VARIANTS", full)
+            .define("GGML_VULKAN", full)
             .define("GGML_BACKEND_DIR", &backends);
     }
     if os == "macos" {