Attached are a PNG and a dat-file - pak them in 128 size, and you'll find the test image as a track.
It is wrong in the first place. There is no sRGB to linear conversion of the base colours before blending and the results of blending are not converted from linear to sRGB. The blend maths only works in linear colour space as opposed to displays which are in the sRGB colour space, or have an sRGB like response to them.
Yeah, yeah, I know how that part works. That explains why the colors would be off, depending on what colors you mix together, especially with bright colors. But there is an issue with this theory:
1) It would not matter which of the color is "background" and which is transparent in the foreground. But with player colors, you only get problems if the player color is transparent, not if you blend something with solid player color.
2) For the blending, it would not matter whether it's player color or a normal color, but these results only come up with player colors.
3) The discoloration would be a result of the blend, meaning it would occur if the background and the foreground don't play along well. But here, the discoloration is always the same, independent of the background.