I guess the idea is that an OsStr is a platform-defined superset of valid UTF-8. This is an arbitrary byte string in the case of Unix, and UTF-8 + surrogate codepoints (also known as WTF-8) in the case of Windows.
If you have a str, call str::encode_utf16() to get the code units. If you have an OsStr on Windows, call OsStrExt::encode_wide() to get the code units. Collect either of these iterators into your fixed-size working buffer of choice.
In fact, RFC 2295 (os_str_pattern) aims to extend the API surface of OsStr, moving its Windows representation from WTF-8 to OMG-WTF-8 in the process. Incidentally, a recent issue regarding OsStr notes the same thing that you do, that &str to &OsStr must always work.
Yes, WTF-8 is a strict superset of UTF-8, so all valid UTF-8 byte sequences are also valid WTF-8.
You know, you might be right there. Slice::from_str() depends on Slice and Wtf8 having the same layout: