> The algorithm isn’t single-pass in the sense of Adaptive Huffman Coding: it still uses the normal Huffman algorithm, but the input is transformed in the same traversal that builds the tree to transform it.

Limited understanding here too. Sounds like it's not really single pass anyway so it's not usable to process a stream in real-time either, before having all the data?

It's a weird claim about a single pass too. It's more of a "let's replace some iteration with building a tree of functions to call" and then pretends waking/executing that is not another pass.

There's no (practical) advantage to the circular implementation; it's just a curiosity.

It is useful for understanding laziness and some interesting theoretical tools for traversing data structures, though. For a more in-depth look at the idea of circular programs for traversal, Bird's paper (linked in the post, https://link.springer.com/article/10.1007/BF00264249) is a good start.