Richard has decided that ".." shall not be used to cross file handler
boundaries.  So file-relative-name needs to be changed to something
like Lars suggested.

I have some comments/questions about this implementation.

Why the extra arg SEPARATE-TREES?  If I understand Richard correctly,
he wants file-relative-name to always behave as if SEPARATE-TREES was
true.

I agree, that parameter should be removed.

>                   (equal
>                     (and
>                       (string-match re filename)
>                       (substring filename 0 (match-end 0)))
>                     (and
>                       (string-match re directory)
>                       (substring directory 0 (match-end 0)))))))))


I'm not sure it's a good idea to check this stuff here.  Maybe it
would be better to let each filename handler do that.

For example, Tramp has the concept of a default method, so
/user@host:/file and /ssh:user@host:/file denote the same file by
default, even though the strings are different.

It seems to me that it is better to let the handler decide.

I don't see a problem. My implementation calls expand-file-name on filename as well as directory. I guess that should expand "/user@host:/file" to e.g. "/ssh:user@host:/file". Am I wrong?