hnswlib stubtest fails for unknown reasons

[github-actions]

Run listed here: https://github.com/python/typeshed/actions/runs/30317471339

[nightcityblade]

Hi, I'd like to investigate and fix this daily test failure. I'll submit a focused PR shortly.

[srittau]

This is a strange one. Third party stubtest fails with an empty output for hnswlib. When running locally, it succeeds.

**********************************************************************

Commands run:

/tmp/stubtest-lvxp8txo/bin/pip install hnswlib[]==0.8.* mypy==2.3.0 numpy>=1.21
MYPYPATH=stubs/hnswlib /tmp/stubtest-lvxp8txo/bin/python -m mypy.stubtest --mypy-config-file /tmp/tmppn1jx5iv --show-traceback --strict-type-check-only --custom-typeshed-dir . hnswlib --allowlist stubs/hnswlib/@tests/stubtest_allowlist.txt

**********************************************************************

Command output:


**********************************************************************

Python version: Python 3.13.14 (main, Jun 11 2026, 03:02:07) [GCC 13.3.0]

Ran with the following environment:
ast_serialize==0.6.0
hnswlib==0.8.0
librt==0.13.0
mypy==2.3.0
mypy_extensions==1.1.0
numpy==2.5.1
pathspec==1.1.1
pip==26.2
typing_extensions==4.16.0

**********************************************************************

To fix "unused allowlist" errors, remove the corresponding entries from stubs/hnswlib/@tests/stubtest_allowlist.txt


**********************************************************************

Upstream repository: https://github.com/nmslib/hnswlib
Typeshed source code: https://github.com/python/typeshed/tree/main/stubs/hnswlib

**********************************************************************

[srittau]

After #16126 this exits with code -4:

hnswlib... (6.45 s) failed with exit code -4

This indicates that the process was killed with SIGILL (an illegal instruction). It's still a bit weird, as hnswlib is distributed source-only, so CI should compile the extension itself.

[JelleZijlstra]

That's weird! The next useful step would be to try to reproduce it and then put gdb on it to figure out what's going on.

But I don't think I'll have time to do that. Disabling stubtest for this library seems like a reasonable step for now.

[srittau]

Using darwin seems to work as well. 🤷‍♂️