> My understanding is that 4 is a relatively sweet spot for allowing efficient consing/deconsing of the finger nodes.
I doubt it. For the persistent data structures I've implemented a branching factor of 32 or 64 worked best (persistent hash maps, vectors, and b+ trees). A high branching factor keeps the depth of the tree very small. Copying a block of memory a few times is cheaper than copying a little memory block a lot of times.