After further study, I realized that the first path to a goal generated may not be the shortest path. It needs to keep generating goals until no new or shorter paths are found. It seems this doesn't happen all that often, so it hardly affects the runtime.
Eliminating some of the low level structure, the full matrix including paths can be saved, at least as high as 256. Judging by the memory usage, it could probably go a bit higher.
Here's the results I have for 256. Hopefully they pass all of plasmid's tests this time.
