As per our discussions on Discord recently, it's probably worth figuring out some additional special treatment for transparent pixels in the Octree Hexadecatree, and in the distance metric. Maybe this would be a follow-up PR at some point. I'm adding this here so there's an easy-ish place to find this info, not because I think this PR needs to incorporate this.

The gist of what I've figured out so far is that when A=0, the distance to any non-A=0 pixel should be the alpha value of that other pixel and should disregard the color channels. So Distance(Rgba(0,0,0,0), Rgba(0,0,0,255)) is 255, but Distance(Rgba(32,64,128,0), Rgba(128,64,32,255)) is also 255. A transparent color is equal to any other transparent color (and this is exactly how it works in premultiplied alpha).

This is important when building the octree so that you don't allocate more than 1 palette slot for fully transparent.

This is also very important during the error diffusion process. If the source color was, say, Rgba(16,16,16,16) but the closest color in the palette was Rgba(0,0,0,0) then the only actual error is that the alpha channel isn't 0. The values of the color channels don't even matter at that point.

We haven't quite figured out the right formula for when working with alpha values other than 0 or 255. My quantization implementation only handles a single A=0 slot in the palette so I haven't had a chance to explore this properly yet. @saucecontrol has a good theory, that the error should be deltaC*(1-abs(deltaA)) but I'm not yet sure it's complete because it doesn't give me good results in some cases (and of course it's possible that the problem is in my code, not necessarily in that formula).