I had a really great brief ability to think of trees normally, and it
   seems the thing to do is to ensure that every included entry in an
   index is a full tree (# leaves = 1 << height) (since new leaves are
   flushed in the same data as the root, "balanced" trees have unbalanced
   shape with extra leaves on internodes), and I think treat each flush as
   if it is building a full tree.
   This could get complicated for random writes, but isn't for just
   appending. I got inhibited when working to accumulate children of last
   non-full tree so as to treat a subregion of it as a full tree. Not
   complex to implement, just inhibited.