@@ -0,0 +1,143 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+"""Contract tests for the host-callback helper.
+
+* how a Python exception raised inside a callback is reported (the trampoline is
+  ``noexcept with gil``, so nothing may cross the C ABI),
+* that a caller-owned raw ``user_data`` pointer reaches a ctypes callback
+  unchanged,
+* that two devices running their own graphs keep callback parameters and results
+  separate.
+"""
+
+import ctypes
+import sys
+from contextlib import ExitStack
+
+import pytest
+
+from cuda.core import Device, LegacyPinnedMemoryResource
+from cuda.core.graph import GraphDefinition
+
+N = 16
+
+
+def _as_bytes(buffer):
+    return bytes((ctypes.c_uint8 * N).from_address(int(buffer.handle)))
+
+
+@pytest.mark.agent_authored(model="gpt-6-astra")
+@pytest.mark.thread_unsafe(reason="replaces the process-global sys.unraisablehook")
+def test_python_callback_exception_is_reported_as_unraisable(init_cuda, monkeypatch):
+    """A raising callback is reported, not propagated, and the graph still finishes."""
+    records = []
+    monkeypatch.setattr(sys, "unraisablehook", records.append)
+    ran = []
+    dev = Device(0)
+    dev.set_current()
+    stream = dev.create_stream()
+
+    def throws():
+        ran.append("callback")
+        raise RuntimeError("callback raised")
+
+    with (
+        dev.allocate(N, stream=stream) as device_buffer,
+        LegacyPinnedMemoryResource().allocate(N) as host_buffer,
+    ):
+        graph_def = GraphDefinition()
+        graph_def.memset(device_buffer, 0x33, N).callback(throws).memcpy(host_buffer, device_buffer, N)
+        graph = graph_def.instantiate()
+        graph.upload(stream)
+        graph.launch(stream)
+        stream.sync()
+
+        later = []
+        after_def = GraphDefinition()
+        after_def.callback(lambda: later.append("later"))
+        after = after_def.instantiate()
+        after.upload(stream)
+        after.launch(stream)
+        stream.sync()
+
+        assert ran == ["callback"]
+        assert [record.exc_type for record in records] == [RuntimeError]
+        assert "callback raised" in str(records[0].exc_value)
+        # The node after the raising callback still ran: the launch and the sync
+        # returned normally instead of turning the exception into a CUDA error.
+        assert _as_bytes(host_buffer)[:4] == b"\x33" * 4
+        assert later == ["later"]
+
+
+@pytest.mark.agent_authored(model="gpt-6-astra")
+def test_ctypes_callback_receives_caller_owned_user_data_pointer(init_cuda):
+    """``user_data`` given as an int is the caller's pointer, passed through as-is."""
+    dev = Device(0)
+    dev.set_current()
+    stream = dev.create_stream()
+    callback_type = ctypes.CFUNCTYPE(None, ctypes.c_void_p)
+    seen = []
+
+    @callback_type
+    def read_byte(data):
+        seen.append((data, ctypes.cast(data, ctypes.POINTER(ctypes.c_uint8))[0]))
+
+    with LegacyPinnedMemoryResource().allocate(4) as payload:
+        bytes_view = (ctypes.c_uint8 * 4).from_address(int(payload.handle))
+        bytes_view[:] = [0x5A, 0x00, 0x00, 0x00]
+        raw_pointer = int(payload.handle)
+
+        graph_def = GraphDefinition()
+        graph_def.callback(read_byte, user_data=raw_pointer)
+        graph = graph_def.instantiate()
+        graph.upload(stream)
+        graph.launch(stream)
+        stream.sync()
+
+        assert seen == [(raw_pointer, 0x5A)]
+        # The caller keeps owning the storage; the graph copied nothing.
+        bytes_view[:] = [0x00, 0x00, 0x00, 0x00]
+
+
+@pytest.mark.agent_authored(model="gpt-6-astra")
+def test_two_devices_keep_callback_parameters_separate(device_x2):
+    """Two devices, two graphs each: no callback observes another one's value."""
+    per_stream = 2
+    slot_count = len(device_x2) * per_stream
+    runs = []
+    streams = []
+    expected = []
+    with ExitStack() as owned:
+        # One pinned allocation, cut into one distinct slot per (device, stream), so
+        # the callbacks cannot alias each other's storage; the owner lives here.
+        host_slots = owned.enter_context(LegacyPinnedMemoryResource().allocate(N * slot_count))
+        host_base = int(host_slots.handle)
+        for device_index, dev in enumerate(device_x2):
+            dev.set_current()
+            for stream_index in range(per_stream):
+                value = 0xA1 + device_index * 0x10 + stream_index
+                slot = device_index * per_stream + stream_index
+                expected.append(((device_index, stream_index), value))
+                device_stream = dev.create_stream()
+                device_buffer = owned.enter_context(dev.allocate(N, stream=device_stream))
+                host_pointer = host_base + slot * N
+
+                def snapshot(label=(device_index, stream_index), host_pointer=host_pointer):
+                    runs.append((label, ctypes.c_uint8.from_address(host_pointer).value))
+
+                graph_def = GraphDefinition()
+                graph_def.memset(device_buffer, value, N).memcpy(
+                    host_pointer, device_buffer, N, dst_owner=host_slots
+                ).callback(snapshot)
+                graph = graph_def.instantiate()
+                graph.upload(device_stream)
+                graph.launch(device_stream)
+                streams.append(device_stream)
+
+        for device_stream in streams:
+            device_stream.sync()
+
+    assert sorted(runs) == sorted(expected)
+    assert len(runs) == slot_count