The reason is because the camera sensor doesn't actually record colors. Instead it records the amount of light received per pixel after some wavelengths are (mostly) rejected (filtered out by the RGB color filter array (CFA)).

During the demosaicing process the color of each pixel is then calculated using the values of the photosite's recorded brightness as well as the values of photosites surrounding it. However, the different pixels/colors are not weighted equally. I.e. the green centric pixels are weighted more heavily in terms of exposure because human perception is also more sensitive to green centric wavelengths. There's also 2x as many green filtered photosites as there are red and blue.

For this reason, significant under/over exposures (exposure shifts) will result in different colors that cannot be recovered in post. I.e. photographing a color chart using the ETTR technique will result in colors that cannot be recovered to the same values as they would record at with a neutral (unshifted) exposure.

The specifics of the demosaicing algorithm vary (as do CFA's, etc); but I'll give a hypothetical example.

Let's say we have red exposed at a value of 1 green exposed at a value of 2 and blue exposed at a value of 3 (1/2/3); and RGB are weighted at factors of 1/3/2 for a 1/6/6 RGB result. Next we record (or exposure shift) the initial values brighter by a value of 2; so we have 3/4/5 weighted the same 1/3/2, for a 3/12/10 result. And no matter what we do with a linear/global adjustment, we will not be able to get the 3/12/10 result back to a 1/6/6 base result.