Setting an empty string to an environment variable deletes the variable on Windows.
So, exists() will return 0 on Windows.

Setting an empty string to an environment variable deletes the variable on Windows.

Is that a Windows issue, or due to how Vim does it?
It appears to work with Neovim, which uses libuv.

For setting: https://github.com/neovim/libuv/blob/c41b31ed3b88aacae1d8498d0f3f6bbc20d56f9a/src/win/util.c#L1456-L1484

But likely more related to how it gets it: https://github.com/neovim/libuv/blob/c41b31ed3b88aacae1d8498d0f3f6bbc20d56f9a/src/win/util.c#L1456-L1484
It uses GetEnvironmentVariableW and checks for ERROR_ENVVAR_NOT_FOUND when it gets an empty value.

Is that a Windows issue, or due to how Vim does it?

It's the behavior of _wputenv. The set command of cmd.exe is also the same. So, this behavior is natural for Windows users.
It seems that SetEnvironmentVariable API can set an empty string, but it doesn't set an environment variable for C runtime library. So we cannot use it. (E.g. gettext depends on it.)

I see. Should we adjust this test and the one in #4751 for win32 behavior then?

I think so.

btw: I've also noted that getenv() might return something that environ() does not contain:

unlet! $VIMRUNTIME
" Fails
echo environ()['VIMRUNTIME']
" Works
echo getenv('VIMRUNTIME')
echo environ()['VIMRUNTIME']

This is due to the side-effect of vim_getenv handling and setting this.

This is not portable, different systems have different ideas about whether an environment variable with an empty value exists or not.
The special handling of $VIMRUNTIME is documented, I thought. Anyway, one should not delete it.
I don't think we should have a test for undefined behavior.

different systems have different ideas about whether an environment variable with an empty value exists or not.
I don't think we should have a test for undefined behavior.

I think Vim is defining it in this case (since Neovim handles it differently on the same system (Windows)), and if the special behavior with win32 is kept, then it would be good to have this tested / documented.
It is not "undefined" really as in "changes randomly".