Until the mythical hashing algorithm is found, why not just allow modders to add their own entries to hash2map.csv as well? It would be trivial to separate the csv's in \hashmap\ by suffix, let's say _hm and _hm2, and everyone has access to both hashing algorithms, right? So just include both hashes for every image you modify, build collisions.csv on the fly, and there won't even be problems with future collisions (unless of course both algorithms produce collisions on the same files).
Honestly, I'm not convinced that we will be able to find a perfect hashing algorithm without sacrificing efficiency, there's just too many super-similar textures in menus, loading screens, etc. Certainly we can find the set of pixels with the greatest variance throughout the entire dataset of images, but even then a second round of hashing will probably be necessary for at least a few images. As long as we can keep the quickly-reloaded images (like the worldmap textures, apparently) in their own bucket we shouldn't have too much of a problem. Or am I missing something?