Seems like it is not trivial to install numpy 1.26 (probably because of Meson) on CygWin (hence, the pinning in the install step). However, since build isolation is used, it will try to reinstall an unpinned version.

The minver test is probably related, although I cannot really find anything specific in the 1.22 release notes that seems related. The PR that introduced the failing test is: #26253

Seems like there are two floating-point related errors with 1.22. Since these are really small, should I just add a bit of tolerance (and a note that these are for NumPy 1..2)?

An option is, I guess, to try 1.22.1 etc to see if this is a regression that was fixed. (Have checked the change log briefly, but not really sure what the underlying problem is anyway...)

I've had to bump tolerance on at least one of those failing tests in Fedora fedora-python@d429c3d (though I cheated a little bit and put it in the non-x86 commit). I think it's likely to be the SIMD optimization issue that we've run into with newer NumPy.

We discussed on the call today and decided to go directly to 1.23, so I've pushed that, and also updated the missed pyproject.toml.