I still think the easiest thing to do would be to make all non-ASCII characters instances of "invalid_character_token", self-delimiting in the same way that operators are.  That would automatically point to exactly the right place in the token stream, and requires zero changes to the error handling code.

I don't have time to look at the code, but I suspect that you could handle this exactly the same way that ? and $ are handled, and maybe even use the same token type.