We are well aware that we should not nest nested named subs. We don't do that. And the code works normally. It's just that all of a sudden the code stops working (well, the perl interpreter stops working it seems), and the eval counter is at roughly the same position each time the code stops working. Given that the code is non-deterministic, what portion of the code that fails is random.
Some of the code snippets that increase the eval counter are well justified, some are perhaps not. But currently, the non-deterministic application is crashing on a regular basis, and the only common factor we've found is that the counter is around 999970000. It seems like a major coincidence.