No no! I'm well aware that you are flattening the actual array only on demand.

And as I said thats not expensive for binary trees.

I'll qualify it as "half-lazy" if you want, b/c of the on-demand.

IMHO, real lazy means no big-copies at all... (see requirements)

Just imagine a non-binary tree like [[1..1e8]] then your algorithm would double the memory needed.

PS: I wouldn't be surprized if your algorithm was faster than hdb's but with the advantage to be able to handle non-binary trees.