[BUG]: Free-threaded CI: coredump bool regression test races on application-wide settings #3225

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct

Type of Bug

Runtime Error

Component

cuda.bindings

Describe the bug

(From GPT 6.1 Sol High)

Both ARM64 free-threaded jobs in CI run 40126604939 failed in test_coredump_bool_uses_single_byte_storage:

• Job 120970829759: Python 3.15t, CUDA 13.4.2, L4
• Job 120970829705: Python 3.14t, CUDA 13.4.2, L4
  Both fail at the same assertion:

driver.coredump_set_attribute_global(driver.CoredumpSettings.LIGHTWEIGHT, False)
assert driver.coredump_get_attribute_global(driver.CoredumpSettings.LIGHTWEIGHT) is False
# AssertionError: assert True is False

Likely cause
PR #3095 changed this test from inspecting a private helper’s storage to round-tripping an application-wide driver setting. The test remains eligible for parallel execution, and free-threaded CI uses --parallel-threads=4, running multiple copies concurrently.
The setter/getter sequence can therefore interleave:

1. Worker A sets LIGHTWEIGHT=False.
2. Worker B sets LIGHTWEIGHT=True.
3. Worker A reads True and fails.
   The same configurations passed immediately after Ship _v2 lowpp API for the driver module #3095, consistent with an intermittent race. Executing the test against simulated, synchronized global storage also reproduces the assertion failure under this interleaving. A direct CUDA reproduction has not yet been performed.
   Suggested fix
   Mark the test pytest.mark.thread_unsafe, with a reason explaining that it modifies application-wide coredump settings. Consider restoring the original setting in finally to avoid leaving state behind.
   The test should pass consistently under free-threaded CI while retaining coverage of boolean attribute handling.

How to Reproduce

Might not be reproducible. You can re-run CI.

Expected behavior

Expected asserts to be equal.

Operating System

No response

nvidia-smi output

No response