So... why are we still having problems with banding in image compression? If anything, gradients should be the easiest things to compress in these images, because the compression algorithms work entirely in the frequency domain. Whatever is introducing banding here is adding more frequency coefficients and making the image bigger and worse at the same time.

Did Google/On2 just not notice that they were crushing every gradient they encode or is are all the common WebP encoders doing some kind of preprocessing pass that crushes gradients and munges luma?