array_scan reports CUDA error 402 during graph capture #2319

Bug description

Capturing a small integer scan with wp.ScopedCapture prints CUDA error 402, followed by a warning about allocation-dependent nodes.

The graph still captures and replays successfully, the scan returns the expected values, and the process exits with status 0. This reproduces in a standalone script without FEM or conditional graph bodies.

Minimal reproduction

import warp as wp

device = "cuda:0"
source = wp.ones(11, dtype=wp.int32, device=device)
result = wp.empty_like(source)

with wp.ScopedCapture(device=device) as capture:
    wp.utils.array_scan(source, result)

wp.capture_launch(capture.graph)
values = result.numpy().tolist()
print("Scan result:", values)
assert values == list(range(1, 12))
print("Result matches expected values")

Current output

Startup output omitted and the source path shortened:

Warp CUDA error 402: attempted introspection would be semantically lossy (in function get_dependent_leaf_nodes, .../warp/native/cuda_util.cpp:436)
Warp warning: wp_free_device_async: failed to get allocation-dependent nodes
Scan result: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11]
Result matches expected values

Expected behavior

Capturing and replaying this scan should return the expected values without reporting CUDA errors or warnings.

Application / Workflow Impact

Found during isolated validation of FEM restriction changes. The standalone reproduction above also triggers the error without FEM. No application-level failure has been established; the observed scan results are correct despite the reported CUDA error.

Environment

• Warp 1.20.0.dev0, source checkout at 37b3a99
• CUDA Toolkit 13.5; driver 595.58.03, reported by Warp as CUDA driver 13.2
• NVIDIA RTX PRO 6000 Blackwell Server Edition, MIG 2g.48gb, sm_120
• Python 3.12.3, Linux x86_64
• CUDA mempool enabled