On a cursory look it seems to construct a distance matrix between all pairs of items, and look for a spanning tree with that distance metric that is "elongated"

On a cursory look it seems to construct a distance matrix between all pairs of items, and look for a spanning tree with that distance metric that is "elongated" - i.e., less branching. Straightforward enough. There seems to be a bit more, involving iterating over different scales and over different distance metrics, but that's not particularly principled or efficient.