Is it just uppercase letters in the basic source character set, or anything considered an uppercase letter in the universal character set after phase 1 transcoding and universal-character-name resolution? Or is there some other definition of uppercase?
I have a slight preference for restricting to just A-Z so that it doesn't require humans or tools to consult the unicode data tables to decide if an identifier is safe to use.
Proposed resolution:
Replace [lex.names]/3.2 with:
Each identifier that contains a double underscore __ or begins with an underscore followed by an uppercase <del>letter</del><ins>nondigit</ins> is reserved to the implementation for any use.
Alternatively we could either create a new grammar production for uppercase nondigits, or just say something like "one of the universal characters in the range 0041-005A (A-Z)"