test: Explicitly mark char as signed in dtype tests

[pganssle-google]

Description

When building in an environment where char is unsigned by default (e.g. due to -funsigned-char).

This explicitly uses signed char in two tests where the sign of the char is important (though only test_dtype_normalized_num fails without this change).

Example failure:

$ g++ --version
g++ (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0

g++ -o pybind11/tests/test_numpy_dtypes.os -c -std=c++20 -fPIC -fvisibility=hidden -O0 -g -Wall -Wextra -Wconversion -Wcast-qual -Wdeprecated -Wundef -Wnon-virtual-dtor -Wunused-result -Werror -funsigned-char -Wpedantic -isystem /usr/include/python3.12 -isystem /usr/include/eigen3 -DPYBIND11_INTERNALS_VERSION=10000000 -DPYBIND11_SMART_HOLDER_PADDING_ON -DPYBIND11_STRICT_ASSERTS_CLASS_HOLDER_VS_TYPE_CASTER_MIX -DPYBIND11_ENABLE_TYPE_CASTER_ODR_GUARD_IF_AVAILABLE -DPYBIND11_TEST_BOOST -Ipybind11/include -Ipybind11/include -Ipybind11/include pybind11/tests/test_numpy_dtypes.cpp

        for a, b in m.test_dtype_normalized_num():
>           assert a == b
E           assert 1 == 2

a          = 1
b          = 2
byteorder  = 'little'
d1         = dtype({'names': ['a', 'b'], 'formats': ['<i4', '<f8'], 'offsets': [1, 10], 'itemsize': 20})
d2         = dtype([('a', '<i4'), ('b', '<f4')])
e          = '<'
expected_chars = ['b', 'h', 'i', 'l', 'q', 'B', ...]
simple_dtype = dtype({'names': ['bool_', 'uint_', 'float_', 'ldbl_'], 'formats': ['?', '<u4', '<f4', '<f16'], 'offsets': [0, 4, 8, 16], 'itemsize': 32})

test_numpy_dtypes.py:195: AssertionError
===================================================== short test summary info =====================================================
SKIPPED [1] test_stl.py:157: no <experimental/optional>
FAILED test_numpy_dtypes.py::test_dtype - assert 1 == 2
============================================ 1 failed, 1162 passed, 1 skipped in 6.04s ============================================

Suggested changelog entry:

Explicitly used ``signed char`` for two numpy dtype tests. As seen when
compiling using ``clang`` on Linux with the ``-funsigned-char`` flag.

[rwgk]

Looks good to me.

@seberg Is there a chance that you could take a quick look at this 2-lines-test-only change? Is that a good way of dealing with the situation?

@pganssle-google It would be great if you could record in the PR description what the platform is that made this change necessary. Simply: OS, compiler version, example compilation command (with all options; doesn't have to be pretty).

[seberg]

Looks right to me, surprising nobody noticed. NumPy has no "char", just signed/unsigned byte a dtype.

[pganssle-google]

@pganssle-google It would be great if you could record in the PR description what the platform is that made this change necessary. Simply: OS, compiler version, example compilation command (with all options; doesn't have to be pretty).

You might know better than me since you used to own pybind11 in Google, but I believe we're using clang, and looking at the build command, I think building with -funsigned-char should trigger this issue.

I will edit the original PR description, let me know (or edit directly if you can?) if you want any other changes.

[rwgk]

You might know better than me since you used to own pybind11 in Google, but I believe we're using clang, and looking at the build command, I think building with -funsigned-char should trigger this issue.

Confirmed locally, thanks! Same with gcc.

We should probably have some builds in our CI using -funsigned-char. I'll try to do that soon.

[rwgk]

I'll merge this now, so that it doesn't become a stumbling block for @nevedaren.

We can add testing with -funsigned-char later.