Why test this at all? It's not as if a website should ever need to render a user's password as text. Is there another use case for excluding this "weird stuff" that I'm not seeing?
You don't need to use IDNA for this, though. There are standards specifically for dealing with Unicode passwords, such as SASLprep (RFC 4013) and PRECIS (RFC 7564).
If I want to use äöüßÄÖÜẞ because I'm confident that I can properly type them on all devices I'll need to type then, then let me. It's not your concern what method of input I'm using.
And maybe, just maybe, using latin characters is actually more of a hassle for a user anyway. (I think the risk of that occoring is low, but still. At the moment, it's a self-fulfilling prophecy that all users have proper method to input atin script available. We simply force them to have one.)
Edit: And the confusion is also possible with just latin characters. U+0430 looks exactly like "a", but has a different code point and thus ruins the hash.