diff --git a/.github/workflows/native-core-build.yml b/.github/workflows/native-core-build.yml
index f201147b..f8959cb4 100644
--- a/.github/workflows/native-core-build.yml
+++ b/.github/workflows/native-core-build.yml
@@ -7,6 +7,9 @@ on:
       # FFI-binding Python files that import the native `_core` module.
       - "agent_assembly/__init__.py"
       - "agent_assembly/types.py"
+      # The modules test_native_core_runtime.py drives through `_core`.
+      - "agent_assembly/core/**"
+      - "test/integration/test_native_core_runtime.py"
       - ".github/workflows/native-core-build.yml"
   workflow_dispatch:
 
@@ -53,3 +56,17 @@ jobs:
       - name: Verify native module import
         run: |
           uv run python -c "from agent_assembly._core import RuntimeClient, GovernanceEvent"
+
+      # AAASM-6359: this file skips itself unless the gate variable is "1", and a
+      # skip exits 0, so a typo'd/removed variable would turn the job into a
+      # silent pass. `-o addopts=""` drops pytest.ini's `--reruns 1` (a retry
+      # would hide a flaky timing assertion) and the coverage plugins; the
+      # junit check fails the job if anything was skipped.
+      - name: Run native core runtime tests
+        env:
+          AAASM_RUN_NATIVE_CORE_TESTS: "1"
+        run: |
+          uv run --with pytest python -m pytest -o addopts="" -rs \
+            --junitxml=native-core-runtime.xml \
+            test/integration/test_native_core_runtime.py
+          grep -q 'skipped="0"' native-core-runtime.xml
diff --git a/test/integration/test_native_core_runtime.py b/test/integration/test_native_core_runtime.py
index e96f23db..402befee 100644
--- a/test/integration/test_native_core_runtime.py
+++ b/test/integration/test_native_core_runtime.py
@@ -177,15 +177,16 @@ def test_send_event_is_non_blocking(native_core: Any) -> None:
     try:
         # Enqueueing 500 events measures a steady ~54 ms on the current shared
         # client (JSON -> wire encoding included), so the original 50 ms budget
-        # was simply below the real cost. A blocking send would wait for the
-        # runtime per event and cost orders of magnitude more, so a 5x margin
-        # still separates "enqueues" from "blocks".
+        # was simply below the real cost. Under CPU contention the same loop
+        # measured 255-286 ms (2 of 14 runs against a 250 ms budget), so the
+        # budget is 1 s: still far below a send that waits on the runtime per
+        # event, and it no longer flakes on a busy runner.
         events = [native_core.GovernanceEvent(make_audit_entry_payload(index)) for index in range(500)]
         start = time.perf_counter()
         for event in events:
             client.send_event(event)
         elapsed_ms = (time.perf_counter() - start) * 1000.0
-        assert elapsed_ms < 250.0, elapsed_ms
+        assert elapsed_ms < 1000.0, elapsed_ms
     finally:
         client.close()
         server.close()