Hammer's arena allocator is, in fact, hot garbage. The entire reason it's there in the first place is that I didn't want to get into the weeds integrating a GC when building a proof of concept that PEGs could work at runtime in C with replaceable backends, so I didn't bother tuning it much beyond "4k sounds like a good number".

To me, the valuable part of hammer was the idea that you could use a friendly PEG-like API and end up with an LALR or GLR parser, and that, I believe, was a resounding sucess.

On the other hand, the generated AST turned out to be even less pleasant to work with than I imagined, and the only part of the implementation that's worth salvaging is the bit reader and writer routines.