Can I revisit this one? @carlkl pointed out that this is a rather significant change in behavior. Previously, for platforms and compilers where double == long double, at Numpy compile time, such as MSVC, this would give npy_longdouble as double. This is useful in the case when we are compiling using Numpy, but with a compiler where it is not true that double == long double. One example is compiling Scipy with mingw-w64, where, by default long double is Float80. Could y'all say more about why this change is needed? Why do we need to force npy_longdouble to be the compiler long double, rather than Numpy long double?

Originally posted by @matthew-brett in #20206 (comment)

Also pinging @seiko2plus, @mattip, and since it seems faintly C++ related @serge-sans-paille