How it should probably work is like a tokenizer. Recognize the longest prefix of the remaining input that can Huffman-decode. Remove that input, and repeat.

Even that won't find the maximal amount of decoding that is possible; for that you have to slide through the input bit by bit and try decoding at every bit position.

However, it seems fair because you wouldn't disassemble that way. If you disassemble some bytes successfully, you skip past those and keep going.