@@ -524,6 +524,8 @@ def record_if_oom(item, call, report, recorder=None, snapshot=None):
     if not recorder.matches(exc_text):
         return None
 
+    # A later native crash can prevent pytest from printing its failure summary.
+    exc_text = f"{exc_text}\n\noriginal traceback:\n{report.longreprtext}"
     text = recorder.capture(item.nodeid, call.when, exc_text, item.config.rootpath, snapshot=snapshot)
     if text is None:
         return None
 
@@ -112,8 +112,8 @@ def test_patterngen_values():
 # driver, and only through the side-effect-free prefix of the checker.
 
 
-def _fake_report(failed):
-    return types.SimpleNamespace(failed=failed)
+def _fake_report(failed, longreprtext=""):
+    return types.SimpleNamespace(failed=failed, longreprtext=longreprtext)
 
 
 def _fake_call(exc_text, when="call"):
@@ -180,6 +180,32 @@ def test_oom_diagnostics_write_an_artifact_naming_the_failing_test(tmp_path, mon
     assert "tests/test_x.py::test_first" in text
 
 
+@pytest.mark.agent_authored(model="gpt-6")
+def test_oom_diagnostics_preserve_original_traceback_immediately(tmp_path):
+    def fail_cuda_sync():
+        raise RuntimeError(OOM_MARKER)
+
+    with pytest.raises(RuntimeError) as excinfo:
+        fail_cuda_sync()
+
+    lines = []
+    item = _fake_item(tmp_path)
+    item.config.pluginmanager.get_plugin = lambda _: types.SimpleNamespace(write_line=lines.append)
+    report = _fake_report(True, longreprtext=str(excinfo.getrepr()))
+    text = record_if_oom(
+        item,
+        types.SimpleNamespace(excinfo=excinfo, when="teardown"),
+        report,
+        recorder=OomDiagnosticsRecorder(),
+        snapshot=_NO_CONTEXT_SNAPSHOT,
+    )
+
+    written = (tmp_path / DEFAULT_FILENAME).read_text(encoding="utf-8")
+    assert "fail_cuda_sync" in written
+    assert report.longreprtext in written
+    assert text in lines
+
+
 @pytest.mark.agent_authored(model="claude-sonnet-5")
 def test_oom_diagnostics_summary_points_at_the_artifact(tmp_path):
     # The report is emitted beside the failing test, thousands of lines above
 
@@ -1275,7 +1275,7 @@ def test_pinned_memory_resource_kernel_access(init_cuda, pool):
 
     n = 1024
     mr = create_pinned_memory_resource_or_xfail(options, xfail_device=device)
-    buffer = _allocate_pinned_buffer_or_xfail(mr, n * ctypes.sizeof(ctypes.c_int), device=device)
+    buffer = mr.allocate(n * ctypes.sizeof(ctypes.c_int), stream=device.default_stream)
     stream = device.default_stream
     stream.sync()
     data = (ctypes.c_int * n).from_address(int(buffer.handle))