I am keeping it as a cache, because the user is allowed to alter the behavior of the parser between parses. And that is useful. I also agree that the cache-dealing code has by now grown out to a horrid state, and I was already playing with the idea of replacing it with a single PV that holds a memcpy of the csv struct (which might need some extending then).

In earlier days, the end user was allowed to alter the hash. I now only allow (reliable) changes through the method calls. That implies that I in theory can move all code to XS.

I however am not yet sure if that would simplify the code, though it probably will :)