I'll see if I can't hack something together using lisp- or scheme-mode with redefined keywords over the weekend. That is, unless someone beats me to it.
I imagine it only officially supports ASCII because it will be migrated away from MzScheme eventually.
Note: Those "u"s are supposed to have umlauts, but that's apparently normalized away somewhere. The point is, u with an umlaut is treated as a single character by the current implementation.
I'm sure it'll be agnostic, if by "agnostic" you mean that it just reads in strings as a sequence of bytes. It would be easier to do that than to check for non-ASCII characters and handle them specially.
I'm not sure PHP's really up to the task - lacking closures might prove to be a big issue. It would probably be reasonably simple to make an arc interpreter, or an interpreter for a simpler intermediate language, but performance would probably suffer.
Why not write an Arc interpreter in Perl -- or even an interpreter for a web-centric subset of Arc? Stick the interpreter in the cgi-bin directory and use that to run your Arc scripts. Perl/CGI is even more ubiquitous on shared hosting accounts than PHP, supports closures, et cetera.
In fact, I've considered creating a new closures-based object system in Perl 5 in the past (never got around to it), which sounds suspiciously like it might have some things in common with some characteristics of Arc.
