It uses capture groups and puts them back in the replacement (the \1 ), so a match of "one" is replaced with (thematch)1(thematch). So eightwo would replace the two with two2two and end up like eightwo2two, so then the t is preserved and eight is found in a later step and then you end up with eight8eightwo2two, and can solve that using part1 only looking for numbers.
For my own I did the same, but without capture groups. So replaced "one" with "one1one" naively without regex and did that for all numbers.
While it maybe is post-hoc clever, I think most of us ended up there because we tried naively to just replace "one" with "1", which broke on things like twone. So when down that path one had to amend it instead of going somewhere else.
