fix(cuda.bindings): make cythonization warning-clean and enable -Werror

[atiaomar1978-hub]

Summary

• Clear the Cython warnings that blocked enabling warning_errors for cuda.bindings (Fixes cuda.bindings: make cythonization warnings clean #2450)
• Drop ignored except * on Python-returning cudla cpdefs and reseal those generated files
• Declare LOAD_LIBRARY_SEARCH_SYSTEM32 as const DWORD in hand-written windll.pxd
• Enable _CythonOptions.warning_errors = True in cuda_bindings/build_hooks.py (matching cuda.core)

Cybind follow-up: the cudla emitter should stop attaching exception clauses to cpdefs that return Python objects so the next sync does not reintroduce those warnings.

Test plan

• Local Cythonize of cuda.bindings with warning_errors=True succeeds (27 extensions)
• toolshed/check_generated_file_seals.py passes for updated cudla.{pxd,pyx}
• CI source builds for cuda.bindings on Linux/Windows

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[atiaomar1978-hub]

Hi maintainers — gentle follow-up on this PR for #2450.

Could someone please:

1. Vet / approve the PR so NVIDIA runners can run CI (copy-pr-bot is still waiting)
2. Review when convenient
3. Assign related issue cuda.bindings: make cythonization warnings clean #2450 to me (@atiaomar1978-hub) so I can own follow-through

Happy to adjust based on feedback (including the cybind sync note in the description).
cc @juenglin (issue author)

[juenglin]

Thanks, Omar.

You may have noticed that some files under cuda_bindings are auto-generated, and the fix for those need to be made in a different place (the generator). We need to wait for that fix before we can merge this PR.

Regarding the pre-commit.ci failure, you might need to run pre-commit install in your virtual environment.

[atiaomar1978-hub]

Thanks @juenglin — that makes sense, and I appreciate the guidance.

On the generated cudla files
Agreed that the lasting fix belongs in the cybind/generator path, not hand-editing sealed outputs. This PR already notes that in the description (the cudla emitter should stop attaching exception clauses to Python-returning cpdefs). I’m happy to hold merge until that generator change lands, then refresh/reseal cudla.{pxd,pyx} from a clean sync so we don’t reintroduce the warnings.

What remains useful in this PR once cybind is fixed:

• windll.pxd: declare LOAD_LIBRARY_SEARCH_SYSTEM32 as const DWORD (hand-written)
• build_hooks.py: enable _CythonOptions.warning_errors = True (matching cuda.core)
• source-level regression tests in test_cython_warning_cleanliness.py

If you’d rather split those from the temporary cudla reseal, I can reshape the PR — just say the word.

On pre-commit.ci
The failure was ruff format (one file). I’ve pushed a signed fix on this branch: a9f3036 (style: ruff-format cython warning cleanliness tests). pre-commit.ci is green again.

Ask
When convenient, could you (or another maintainer) assign #2450 to me (@atiaomar1978-hub) and vet the PR for NVIDIA runners so full CI can run once we’re unblocked on cybind? Happy to adjust based on your preferred sequencing.

[mdboom]

This looks great.

I think we should remove the test_cython_warning_cleanliness.py test. It just adds test complexity we need to maintain, and with "warnings as errors" turned on in build_hooks.py we will automatically catch these warnings (or any others) if they creep back.

[atiaomar1978-hub]

Thanks @mdboom — agreed. I've removed cuda_bindings/tests/test_cython_warning_cleanliness.py in 1695c51; warning_errors in build_hooks.py is enough to catch any regression. PR description/test plan updated accordingly.

[mdboom]

Thanks, @atiaomar1978-hub!

[mdboom]

/ok to test 1695c51

[atiaomar1978-hub]

Hi @mdboom,

Thank you for the review and approval — much appreciated. I've removed the source-level cleanliness tests as requested, and I'm glad the remaining approach (enabling warning_errors in build_hooks.py) looks good.

I also checked the current CI failures on this PR; none appear to come from the PR changes:

• Build win-64 / py3.11 — cibuildwheel failed while downloading virtualenv.pyz with http.client.RemoteDisconnected (network/infra flake). Other Windows Python versions on the same matrix passed.
• Docs / Build docs — the docs build itself succeeded; deploy to gh-pages was rejected by a repository rule (Cannot force-push to this branch).
• PR has assignee, labels, and milestone — process metadata check (assignee/milestone). Label cuda.bindings is already present.
• Check job status — aggregate of the above.

pre-commit.ci, Linux builds/tests, and the other Windows builds look green. If convenient, could you please re-run the flaky Windows job (and set assignee/milestone if needed for the metadata check)? Happy to adjust anything else on my side.

Thanks again for your help on this.

[mdboom]

We are having merge conflicts in generated files here. The easiest thing will be to wait for #2603 to merge and then remove the cudla changes here. I can handle getting that across the finish line. Thanks again for your work, @atiaomar1978-hub.

[mdboom]

/ok to test ce9fb70

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-2463/
https://nvidia.github.io/cuda-python/pr-preview/pr-2463/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-2463/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-2463/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[atiaomar1978-hub]

Thanks @mdboom