Also, they seemingly already had a BWT lying around (which is seemingly not counted in “from scratch in a couple of days”), and there is no mention of speed.

He didn't specify that the demo of BWT was in Ada, or his own.

It seems like author means his implementation is competitive in terms of _compression ratio_ only, and this is achieved in the final Huffman encoding step.

From that perspective the BWT implementation isn't very interesting because the BWT is deterministic, so the effect on compression ratio doesn't depend on which implementation was used.

The data in question cannot be compressed to zero, or close to zero, by far.