CI: Random segfaults on Windows

[leofang]

https://github.com/NVIDIA/cuda-python/actions/runs/37536403033/job/112523008786?pr=3037#step:32:1892

[leofang]

Scanning the CI workflow runs between 2026-09-29 and 2026-10-06, I found 5 more Windows access violation (STATUS_ACCESS_VIOLATION, 0xC0000005) instances matching this signature, every one on free-threaded Python 3.14t + a100 MCDM:

Date UTC	Context	SHA	CUDA	Crashing test	Job
2026-10-02 16:57	#2947	e2ae412	13.4.2 wheels	test_linker.py::test_incremental_ltoir_round_trip	log
2026-10-03 03:47	#3013	b8b2ffe	13.0.2 wheels	test_linker.py::test_incremental_ltoir_round_trip	log
2026-10-05 00:38	scheduled on main	ab9e20f	13.0.2 wheels	test_memory.py::test_device_memory_resource_with_options	log
2026-10-06 17:24	#3028	ef9b712	13.4.2 wheels	test_linker.py::test_incremental_ltoir_round_trip	log
2026-10-06 22:14	#3037 (this issue)	8deb152	13.4.2 wheels	test_linker.py::test_incremental_ltoir_round_trip	log

Observations:

• Every hit is win-64 + Python 3.14t + a100 MCDM. Zero access violation on 3.10–3.14 (non-free-threaded) or 3.15 across all 48 Windows Run cuda.core tests failures in the window — free-threading looks necessary to trigger.
• 4/5 crash in test_incremental_ltoir_round_trip — strongest reproducer candidate.

For reference, in the same window two additional Windows failures surfaced Windows fatal exception: code 0xC0000374 (STATUS_HEAP_CORRUPTION, exit 127) rather than access violation + 139 — different code, but same environment (win-64 + 3.14t + a100 MCDM), so a shared memory-safety root cause is plausible:

• 2026-09-30 — build: consolidate toolchain helpers via symlink #2965 → log, crash in test_linker.py::test_incremental_lto_cubin_round_trip
• 2026-10-02 — ci: keep test failures visible in GitHub Actions logs #2963 → log, crash in graph/test_graph_definition_lifetime.py::test_destroying_conditional_node_invalidates_branch_handles

-- Leo's bot

[leofang]

Correction to my earlier comment: the sweep filtered by step-number instead of step-name and missed 2 instances where "Run cuda.core tests" was numbered 31 (not 32). Both newly surfaced hits are Python 3.13 (non-free-threaded) + rtxpro6000 MCDM, so free-threading is not required to trigger — contrary to what I wrote above.

Date UTC	Py	Context	SHA	CUDA / GPU	Crashing test	Job
2026-09-29 20:33	3.13	scheduled on main	e8d9075	13.0.2 wheels / rtxpro6000 MCDM	test_green_context.py::test_memory_node_updates_preserve_green_context (crash in the green_ctx fixture)	log
2026-10-01 19:56	3.13	#2988	144e356	13.0.2 wheels / rtxpro6000 MCDM	test_green_context.py::test_memory_node_updates_preserve_green_context (same green_ctx fixture)	log

-- Leo's bot