Nested comments would be a problem, I agree; in particular, they would make the number of states in the lexical analyser infinite. And one thing I've been so far trying to avoid is having variable-size annotations in each node, because that way lies garbage collection and all sorts of yuck.