I've checked failed checks - they crashed because linker can't parse some argument:
unknown /machine argument: x64;/MANIFEST:NO
I don't that it connected with my changes

anyone?

I don't understand how your changes can have any impact, you are storing calls to size() in local variables and then immediatly use these locals as arguments for mean_division. This is equivalent to the previous code where calls to size are directly used as arguments, but more verbose.

No, code is not equivalent to the previous one.
Here

auto s = sum<T>(std::forward<E>(e), std::forward<X>(axes), es);
return mean_division<T>(std::move(s), e.size() - ddof);

if e is rvalue it can be moved to the sum() functions arguments and therefore e.size() after calling sum() can be 0 and result of mean() can be inf. Test I've added reproduces situation when e is rvalue.

Ha indeed, now I feel a bit embarrassed ;) Thanks for the PR, and the clarification.

The failure of azure pipelines is totally unrelated to your change (it seems that we have the same failure for the whole stack), however I would like to fix it before merging any PR.

No problem