These tests seem to test process global cleanup (not sure I follow)

Author: seberg

cuda_core/tests/graph/test_graph_definition_lifetime.py

@@ -592,6 +592,7 @@ def test_event_survives_graph_clone_and_execution(init_cuda):
 # =============================================================================
 
 
+@pytest.mark.thread_unsafe(reason="coalescing and drain thread are process-global properties")
 @pytest.mark.agent_authored(model="gpt-5.6")
 def test_user_object_cleanup_is_coalesced_on_python_thread(init_cuda):
     """More than 32 CUDA callbacks drain through one main-thread pending call."""
@@ -610,6 +611,7 @@ def test_user_object_cleanup_is_coalesced_on_python_thread(init_cuda):
     assert set(finalized_threads) == {main_thread}
 
 
+@pytest.mark.thread_unsafe(reason="scenario needs a whole interpreter, so one subprocess per run suffices")
 @pytest.mark.agent_authored(model="gpt-5.6")
 def test_pending_call_queue_saturation_preserves_cleanup(tmp_path):
     """A full CPython queue neither strands nor mis-threads cleanup."""
@@ -1319,6 +1321,7 @@ def test_memcpy_buffer_survives_close(init_cuda):
     assert list(out) == [0xCD] * 4
 
 
+@pytest.mark.thread_unsafe(reason="release is drained by the process-global cleanup queue, not by this thread")
 @pytest.mark.agent_authored(model="claude-opus-4.8")
 def test_memcpy_buffer_allocations_released_after_graph_destroyed(init_cuda):
     """Destroying the graph frees both memcpy operand allocations.

cuda_core/tests/graph/test_graph_update.py

@@ -196,6 +196,7 @@ def new_callback():
     assert called == ["new"]
 
 
+@pytest.mark.thread_unsafe(reason="release is drained by the process-global cleanup queue, not by this thread")
 @pytest.mark.agent_authored(model="gpt-5.6")
 def test_failed_graph_update_does_not_adopt_attachments(init_cuda):
     """A rejected source graph keeps ownership separate from the exec."""