Spike sometimes crash on delete

[vgene]

Test test_tensor_lifecycle_on_delete sometimes crash the process

[vgene]

The intermittent SIGABRT crashes in the spike test suite only occur under pytest -n auto (64 xdist workers) and never in isolation. They manifest in two different NRT cleanup paths — nrt_tensor_free() during tensor destruction and nrt_close() during Spike.close() — both involving tests that cycle through multiple nrt_init/nrt_close sequences.

We verified that nanobind's keep_alive mechanism correctly destroys the C++ tensor before releasing the Spike reference, so the destruction order is not the issue.

The NRT C library's nrt_tensor_free and nrt_close() itself can abort internally when assertion fail.

The most likely explanation is that nrt_close() performs some internal cleanup (thread shutdown etc) and returns before it's fully complete; under heavy system load from 64 workers, the next nrt_init() starts before the previous cleanup finishes, leaving the NRT runtime in an inconsistent state that triggers an assertion on the subsequent close or tensor free.

Pending further investigation.