I know that it really doesn't address your topic, but I think the reason that you're running into the problem is that it's the wrong approach. An impressive (ab)use of regular expressions, though.
What I've seen of predicate systems leads me to a database approach rather than regular expressions. Given the right data structure, you need to write two parsers: one to convert $facts into rows in tables, and one to convert your queries into SQL. With the right data structure, this is straightforward.
Most of this can be done with three tables: One to contain people, things and attributes (Merlyn, Ovid, Kudra, books, gold, valuable), one to contain verbs ( owns, is ), and one to contain tuples of person/thing/attribute, verb, person/thing/attribute. This leads to something like:
This query, with proper data, can answer the question of 'who owns something valuable' in a very reasonable amount of time. You still do code generation, but now it's SQL instead of RE.
Spring: Forces, Coiled Again!
In reply to Re: Being Forced to Fork with Nested Regular Expressions