One-pass parser
Our current wikitext parser goes through many passes, with many regexps and a few explode()/implode()s. Not only is it kinda slow, it's prone to horrible frightening bugs when different levels interfere with each other (such as the URL-in-URL bugs).
One, or at least fewer, passes would be good.
Magnus has, according to rumour, written such a thing for WINOR. Could it be adapted?