Add Windows arm tests

[K20shores]

Description

In response to the conversation about #5883 in #5919, I'm adding tests for Windows ARM runners, both with and without Mingw. This uses the windows-arm-11 runner.

Fortunately, I didn't see the same build error reported in #5883. These are still good tests to add, though. In my testing, the issue was present in pybind11 3.0.0, but it was fixed when updating to 3.0.1. It appears that #5757 includes a fix that already addresses #5883.

For mingw under windows arm with clang, I've had to update the visibility tests. I believe it's scoped appropriately, but I may be wrong.

Suggested changelog entry:

• Add CI tests for windows-11-arm with clang/MSVC (currently python 3.13), windows-11-arm with clang/mingw (currently python 3.12)

---

📚 Documentation preview 📚: https://pybind11--5932.org.readthedocs.build/

[K20shores]

@rwgk @baburton I've added this PR in response to comments on #5919. It seems the issue has already been fixed in the latest version of pybind11 (3.0.1). At least, it fixes it for me. @baburton, if you try 3.0.1 do you see the same error?

[rwgk]

I worked on this PR with the help of Cursor. Initially I wasn't clear about the nature of the differences between the two new jobs.

After adding two commits, I asked Cursor for it's opinion, this came back (and I agree):

This looks merge-ready to me.

• CI changes: The two Windows ARM jobs are clearly named (windows_arm_clang_msvc / windows_arm_clang_msys2), structurally aligned, and trimmed to a single Python each to respect free GHA limits while keeping the matrix shape extensible.

• Tests/requirements: The NumPy pinning now cleanly separates ARM64 CPython (2.3.0 for ≥3.11) from non‑ARM64, which matches the practical wheel availability story.

• XFAIL: The unnamed‑namespace XFAIL condition now correctly covers “any libc++” (including Windows ARM MSYS2) plus non‑Windows Clang, which aligns with the environments where the issue is known.

@K20shores could you please review the two commits I added (I also merged master)?

• commit 832a6ab — Clarify Windows ARM clang job naming
• commit 40a9321 — Reduce Windows ARM clang matrix size

@baburton could you please confirm that your original issue is resolved already on master (or 3.0.1)?

[K20shores]

@rwgk I left a comment on commit 40a9321. Everything looks good to me, unless you think my comment there requires more changes

[rwgk]

@rwgk I left a comment on commit 40a9321. Everything looks good to me, unless you think my comment there requires more changes

I added commit c727eef

As soon as I see that the windows_arm_clang_msys2 job succeeds I'l merge this PR. Thanks!

[rwgk]

Huh, what's that error?

  error: clangarm64: signature from "Christoph Reiter (MSYS2 development key) <[email protected]>" is invalid
  error: failed to synchronize all databases (invalid or corrupted database (PGP signature))
  Error: The process 'C:\Windows\system32\cmd.exe' failed with exit code 1

[rwgk]

Passed already: https://github.com/pybind/pybind11/actions/runs/20188840738/job/57963304508

[baburton]

@baburton could you please confirm that your original issue is resolved already on master (or 3.0.1)?

I'm out of town for the weekend and not somewhere that I can readily test this; however, looking at the relevant change in the sources between 3.0.0 and 3.0.1 (pybind11 no longer switches on hidden visibility under windows), I would fully expect that this resolves the problem. (I experienced my build failure under 3.0.0, not 3.0.1.) I should be able to test this properly next week.