This was found to cause a build failure in Google's monorepo. It's unlikely to cause issues elsewhere, since you need to have a particular combination of includes from a C++ build to notice the problem. Still, we may as well fix it.

Harmless enough and I don't mind just merging. Is that really a best-practice -- considering Python is a C-API -- to rely on them having the extern wrapping?

Is that really a best-practice -- considering Python is a C-API -- to rely on them having the extern wrapping?

The Python C API docs say they include the appropriate extern "C" statements:

C++ users should note that although the API is defined entirely using C, the header files properly declare the entry points to be extern "C". As a result, there is no need to do anything special to use the API from C++.

Is that really a best-practice -- considering Python is a C-API -- to rely on them having the extern wrapping?

The Python C API docs say they include the appropriate extern "c" statements:

C++ users should note that although the API is defined entirely using C, the header files properly declare the entry points to be extern "C". As a result, there is no need to do anything special to use the API from C++.

(As it happens, they did not. But I sent a PR fixing that in CPython, see above.)

Not sure I am convinced yet that it is a fix, but since NumPy is happy, I don't mind just putting it in :).

Do you need a backport of this because the real fix in Python is slower to backport?

Do you need a backport of this

Seems reasonable. NumPy 2.2 supports 3.10 - 3.13, that is a lot of Python releases to rely on.