TFA's implementation is using a loop over the window size to compare each character to all the other characters, to see if they're unique. So, the article's O(N * W²) is correct, but I agree, the window size is fixed and small; it's fair to boil it down to just O(N).

(I never even considered doing a for loop like that to determine if the window is unique; I just histogram the window, and then check if all the values in the histogram are 1. If yes, unique. If I had been cleverer, I wouldn't rebuild the histogram as the window shifts, but there was no need for that. This is essentially the next solution TFA presents.)