Thanks for the suggestion. I've tried something essentially similar. The problem is one of scaling, and the fact that the real problem doesn't use chars and strings (see Update 1).

I'm not sure you need the lookahead, but I'm more worried about how it affects execution time. Maybe I'll benchmark something on that while I'm taking a break from this problem ;)

Update: I see that the lookahead captures overlaps, while the other doesn't.