feat: social publishing + NuGet #r + move perf + mesh stability batch

.github/workflows/dotnet-test.yml

@@ -13,7 +13,12 @@ on:
 jobs:
   build:
     name: Build and Run Unit tests
-    timeout-minutes: 30
+    # 60 min: setup+restore+build runs ~3 min; tests currently take ~35-40 min
+    # (FutuRe ~6m, Autocomplete ~4m, AI ~3m, Acme ~3m, plus ~30 smaller projects).
+    # Headroom for a single hung project burning its 6m wallclock cap without
+    # truncating the suite. If this gets tight again, the next move is matrix-
+    # parallelizing the per-project loop instead of bumping the cap further.
+    timeout-minutes: 60
     runs-on: ubuntu-latest
 
     steps:
@@ -34,20 +39,110 @@ jobs:
         DOTNET_ENVIRONMENT: Development
         Logging__LogLevel__Default: Warning
       run: |
-        find test -name '*.csproj' \
-          ! -path '*PostgreSql*' \
-          ! -path '*Cosmos*' \
-          ! -path '*Orleans*' \
-          ! -path '*Acme*' \
-          ! -path '*FutuRe*' \
-          -exec dotnet test {} --no-build --verbosity normal -l:trx \
-                --blame-hang-timeout 3m --blame-hang-dump-type mini \; 2>&1 | tee test/test-results.log
+        # Only Cosmos is excluded — needs the Cosmos emulator which is heavy to
+        # set up on a runner. Everything else runs (PostgreSql via Testcontainers
+        # using pre-installed Docker, Orleans via in-proc TestCluster, Acme/FutuRe
+        # via dynamic Code-piece compilation).
+        #
+        # set +e so a hung/aborted test host (non-zero exit) doesn't terminate
+        # the loop — we want EVERY csproj to attempt to run, then surface what
+        # broke after the loop completes. Without this, pipefail propagates a
+        # single hang into a hard SIGTERM that kills sibling test projects and
+        # the artifact-upload step has nothing to publish.
+        #
+        # Disk hygiene: previous "verbose normal | tee -a test-results.log"
+        # form was bloating one log file with every test's structured trace
+        # output. With ~30 csproj × 100s of tests × verbose Trace logs from
+        # MessageHub, the file grew faster than the 14 GB ephemeral disk could
+        # absorb and the runner sent SIGTERM (exit 143) mid-run. Switch to
+        # `--verbosity minimal`, drop the tee, and only persist the per-project
+        # exit markers — trx files still capture full per-test results.
+        set +e
+        : > test/test-results.log
+        echo "::group::Per-project test runs"
+        # `! -path '*/bin/*'`: the build copies sample csproj fixtures into test
+        # bin dirs (e.g. MeshWeaver.Samples.Graph.csproj under
+        # test/MeshWeaver.Acme.Test/bin/Debug/net10.0/SamplesGraph/). They match
+        # the *.csproj pattern but aren't test projects — invoking dotnet test on
+        # each wastes ~1 s × 23 entries per run.
+        #
+        for csproj in $(find test -name '*.csproj' \
+            ! -path '*Cosmos*' \
+            ! -path '*/bin/*' \
+            | sort); do
+          name=$(basename "$csproj" .csproj)
+          echo "::endgroup::"
+          echo "::group::$name"
+          # `timeout 6m`: per-project wall-clock cap. blame-hang-timeout 30s only
+          # fires when xUnit considers a *test* to have stalled — fixture-init
+          # and between-class hangs slip past it. 6 min covers the slowest
+          # legitimate project today (FutuRe ~5 min) with headroom; defense-in-
+          # depth backstop in case any of the skipped Linux-hang projects gets
+          # re-enabled or another regresses to the same pattern.
+          # blame-hang-timeout: 90 s of test-runner inactivity before blame
+          # collects a hang dump and aborts the test host. The previous 30 s
+          # was firing on legitimate teardown — multiple projects (Hosting.Blazor,
+          # Hosting.Orleans, Autocomplete, NodeOperations) showed all tests
+          # passing in <5 s but the testhost taking 30+ s in finalizer/dispose
+          # cleanup of hosted hubs / static caches, then getting killed and
+          # reported as "Test host process crashed" exit=1. 90 s gives genuine
+          # teardown enough room without masking real hangs (the per-project
+          # 6 m wall-clock cap below is the hard backstop).
+          timeout --signal=TERM --kill-after=30s 6m \
+            dotnet test "$csproj" --no-build --verbosity minimal -l:trx \
+              --blame-hang-timeout 90s --blame-hang-dump-type mini
+          rc=$?
+          if [ "$rc" = "124" ] || [ "$rc" = "137" ]; then
+            marker="[CI] $name exit=$rc TIMEOUT (6m wall-clock cap hit — likely fixture/init hang)"
+          else
+            marker="[CI] $name exit=$rc"
+          fi
+          echo "$marker"
+          echo "$marker" >> test/test-results.log
+        done
+        echo "::endgroup::"
+        exit 0
+    - name: Fail on hang / aborted test run
+      # blame-hang aborts the test host and writes a minidump (blame-*.dmp).
+      # Test *failures* are surfaced separately by Publish Test Results below;
+      # this step's only job is to fail loudly on a HANG so it doesn't get
+      # masked behind a "test results published" green checkmark when the trx
+      # contains only what completed before the abort.
+      run: |
+        hang_dumps=$(find test -name 'blame-*.dmp' 2>/dev/null | wc -l)
+        if [ "$hang_dumps" -gt 0 ]; then
+          echo "::error::A test process hung — blame-*.dmp files were written."
+          echo "--- Hang dumps ---"
+          find test -name 'blame-*.dmp' 2>/dev/null
+          echo "--- Per-project exit codes (non-zero indicates the hung project) ---"
+          cat test/test-results.log
+          exit 1
+        fi
     - name: Collect test logs for artifact
       if: always()
       run: |
         mkdir -p collected-logs
         # Find all test-logs directories and collect their contents without renaming
         find . -path "*/bin/*/test-logs/*.log" -type f -exec cp {} collected-logs/ \; 2>/dev/null || true
+        # The MonolithMeshTestBase phase trace + dispose trace live in the
+        # process tempdir (Path.GetTempPath() == /tmp on the runner). They
+        # carry the per-class INIT_MEM / DISPOSE_MEM lines that pinpoint
+        # which test class drives the AI.Test / Autocomplete.Test / Content.Test
+        # / Hosting.Monolith.Test OOM. Copy with rename so they don't collide.
+        if [ -f /tmp/meshweaver-test-trace.log ]; then
+          cp /tmp/meshweaver-test-trace.log collected-logs/_meshweaver-test-trace.log
+        fi
+        if [ -f /tmp/meshweaver-dispose-trace.log ]; then
+          cp /tmp/meshweaver-dispose-trace.log collected-logs/_meshweaver-dispose-trace.log
+        fi
+        # Per-class INIT → DISPOSE memory delta summary (one line per test instance):
+        # managed=… rss=… rssAnon=… unmanaged=… shared=0|1
+        # rssAnon (Linux only) is where the Autofac Reflection.Emit factory pin lives;
+        # unmanaged = rss − managed is the portable approximation. Grep this file to
+        # find the worst leakers without wading through the full per-event trace.
+        if [ -f /tmp/meshweaver-memory-delta.log ]; then
+          cp /tmp/meshweaver-memory-delta.log collected-logs/_meshweaver-memory-delta.log
+        fi
     - name: Publish Test Results
 #      uses: EnricoMi/publish-unit-test-result-action/composite@v2
       uses: EnricoMi/publish-unit-test-result-action@v2.12.0