[eru]

> But shouldn't Huffman coding already detect that same predictability and compress it the same?

Huffman coding only works on individual letters. It doesn't know anything about relationships between adjacent letters.