Add arch to JDK cache path; skip canceled coverage

Include the runner architecture in the JDK extraction root (jdk_root) to avoid cross-architecture temp/cache collisions. Change the coverage-aggregate job condition from always() to ${{!cancelled()}} so aggregation is skipped for cancelled runs, reducing wasted work and noisy outputs.

.github/actions/java/action.yml

@@ -103,7 +103,7 @@ runs:
             jdk_archive="OpenJDK25U-jdk_aarch64_mac_hotspot_25.0.3_9.tar.gz"
             jdk_sha256="7baab4d69a15554e119b86ff78d40e3fdc28819b5b322955c913cebfe3f6a37c"
           fi
-          jdk_root="${RUNNER_TEMP}/stirling-temurin-${jdk_version}"
+          jdk_root="${RUNNER_TEMP}/stirling-temurin-${jdk_version}-${architecture}"
           archive_path="${jdk_root}/${jdk_archive}"
           mkdir -p "${jdk_root}"
           if [ ! -f "${archive_path}" ]; then

.github/workflows/build.yml

@@ -255,7 +255,7 @@ jobs:
   # for whatever did record. Advisory only - intentionally NOT in
   # all-checks-passed, so a flaky aggregate run never blocks merging.
   coverage-aggregate:
-    if: always()
+    if: ${{!cancelled()}}
     needs:
       - build
       - playwright-e2e-live