ci: run status check always but considered cancellations failures

Author: cpcloud

.github/workflows/ci.yml

@@ -24,7 +24,7 @@ jobs:
       CUDA_PREV_BUILD_VER: ${{ steps.get-vars.outputs.cuda_prev_build_ver }}
     steps:
       - name: Checkout repository
-        uses: actions/checkout@08c6903cd8c0fde910a37f88322edcfb5dd907a8  # v5.0.0
+        uses: actions/checkout@08c6903cd8c0fde910a37f88322edcfb5dd907a8 # v5.0.0
         with:
           fetch-depth: 0
       - name: Get CUDA build versions
@@ -42,7 +42,7 @@ jobs:
       skip: ${{ steps.get-should-skip.outputs.skip }}
     steps:
       - name: Checkout repository
-        uses: actions/checkout@08c6903cd8c0fde910a37f88322edcfb5dd907a8  # v5.0.0
+        uses: actions/checkout@08c6903cd8c0fde910a37f88322edcfb5dd907a8 # v5.0.0
       - name: Compute whether to skip builds and tests
         id: get-should-skip
         env:
@@ -123,7 +123,7 @@ jobs:
     name: Test ${{ matrix.host-platform }}
     if: ${{ github.repository_owner == 'nvidia' }}
     permissions:
-      contents: read  # This is required for actions/checkout
+      contents: read # This is required for actions/checkout
     needs:
       - ci-vars
       - build-linux-64
@@ -144,7 +144,7 @@ jobs:
     name: Test ${{ matrix.host-platform }}
     if: ${{ github.repository_owner == 'nvidia' }}
     permissions:
-      contents: read  # This is required for actions/checkout
+      contents: read # This is required for actions/checkout
     needs:
       - ci-vars
       - build-linux-aarch64
@@ -164,7 +164,7 @@ jobs:
     name: Test ${{ matrix.host-platform }}
     if: ${{ github.repository_owner == 'nvidia' }}
     permissions:
-      contents: read  # This is required for actions/checkout
+      contents: read # This is required for actions/checkout
     needs:
       - ci-vars
       - build-windows
@@ -192,7 +192,7 @@ jobs:
       build-ctk-ver: ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
 
   checks:
-    name: Check job status
+    name: Check job status (not cancelled)
     if: always()
     runs-on: ubuntu-latest
     needs:
@@ -202,4 +202,28 @@ jobs:
       - doc
     steps:
       - name: Exit
-        run: exit 0
+        run: |
+          # if any dependencies were cancelled, that's a failure
+          #
+          # see https://docs.github.com/en/actions/reference/workflows-and-actions/expressions#always
+          # and https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/collaborating-on-repositories-with-code-quality-features/troubleshooting-required-status-checks#handling-skipped-but-required-checks
+          # for why this cannot be encoded in the job-level `if:` field
+          #
+          # TL; DR: `$REASONS`
+          #
+          # The intersection of skipped-as-success and required status checks
+          # creates a scenario where if you DON'T `always()` run this job, the
+          # status check UI will block merging and if you DO `always()` run and
+          # a dependency is _cancelled_ (due to a critical failure, which is
+          # somehow not considered a failure ¯\_(ツ)_/¯) then the critically
+          # failing job(s) will timeout causing a cancellation here and the
+          # build to succeed which we don't want (originally this was just
+          # 'exit 0')
+          if ${{ needs.test-linux-64.result == 'cancelled' ||
+                 needs.test-linux-aarch64.result == 'cancelled' ||
+                 needs.test-windows.result == 'cancelled' ||
+                 needs.doc.result == 'cancelled' }}; then
+            exit 1
+          else
+            exit 0
+          fi