Logged In: YES 
user_id=7887

Hummm.. interesting. Indeed, we could use two different
strategies. We could either restore marks whenever failling,
so that backtrackings don't have to worry about that (it
knows it was restored if the recursion failed), or use your
approach and protect only when *not* failling. While I was
thinking about the first one, the later (your suggestion)
seems more logic, and fast.

Either way, we clearly have a mix of both today (for
example, we have no mark protection in OP_REPEAT), and
that's not very smart. I'll study that code further, to make
sure that we're really on the right track, and also check
your bugs.

Thank you very much for explaining and discussing.