Doesn't that example show that a 256-step gradient takes 256 bytes? A 256-step run of the same colour takes just a couple of bytes due to RLE, doesn't it? (Not an expert.)
I debated about adding "after the first row", but didn't (because it doesn't make sense for gradients in the other direction). Either way, both my and Wikipedia's point is that a simple gradient filling a 2D surface costs only a "1D amount" of bytes, which is going to beat JPEG (at "can still read text" levels of quality).