GH-4098: keep the evidence when a job is cancelled at its cap

.github/workflows/dotnet.yml

@@ -41,6 +41,36 @@ jobs:
           CI_JOB_NAME: dotnet-ci
         run: ./build/run-with-cancellation-relay.sh ci
 
+      # GH-4098. This job wedged once and was cancelled at its cap, and every piece of evidence went
+      # with it: a cancelled job's logs are not truncated, they are DISCARDED (BlobNotFound), leaving
+      # only per-step timestamps in the API. The supervisor already writes a partial ledger inside the
+      # runner's grace window when it is cancelled (build/SupervisedTests.cs registerCancellationCapture,
+      # build/StallCapture.cs) -- naming the tests that were still in flight -- but this workflow never
+      # uploaded it, so the one artifact that survives a discarded log was being thrown away with the
+      # runner. tests.yml has had this since GH-3787; dotnet.yml is the workflow that actually wedged.
+      #
+      # `always()` and not `failure()`: a cancelled job is neither success nor failure, and cancellation
+      # is precisely the case this exists for.
+      - name: Upload retry ledger
+        if: always()
+        uses: actions/upload-artifact@v7
+        with:
+          name: test-ledger-dotnet-ci
+          path: artifacts/test-ledger/
+          if-no-files-found: ignore
+          retention-days: 30
+
+      # Only reachable when the runner itself survived. If the OOM killer took a *test* process rather
+      # than the runner service, the kill is recorded here and nowhere else. Same `always()` reasoning
+      # as the upload above.
+      - name: Kernel OOM report
+        if: always()
+        run: |
+          echo "::group::dmesg (OOM / kill events)"
+          sudo dmesg | grep -iE "out of memory|oom-kill|killed process" || echo "no OOM events in dmesg"
+          echo "::endgroup::"
+          free -m
+
       # The CI target now also runs MessageRoutingTests, which boots a real
       # RabbitMQ broker (started inline by the build target via docker compose).
       # Tear it down on exit so the runner is clean for downstream steps and

.github/workflows/tests.yml

@@ -138,8 +138,12 @@ jobs:
 
       # Only reachable when the runner itself survived. If the OOM killer took a *test* process
       # rather than the runner service, the kill is recorded here and nowhere else.
+      #
+      # GH-4098: `always()`, not `failure()`. A job cancelled at the 20-minute cap is neither success
+      # nor failure, so `failure()` skipped this step on exactly the wedges it was written for --
+      # #4083's CIMarten cancellations among them.
       - name: Kernel OOM report
-        if: failure()
+        if: always()
         run: |
           echo "::group::dmesg (OOM / kill events)"
           sudo dmesg | grep -iE "out of memory|oom-kill|killed process" || echo "no OOM events in dmesg"