Relative rather than absolute calibration. For CCT and Duv we just need X, Y and Z to be in the correct proportions given an illuminant.
Looking at the idealized plot in the OP you can see that some ‘color’ channels are more sensitive than others, so that if we actually had an illuminant that supplied the same power per small wavelength interval (Se) we would not get the same count per channel out of the chip. In addition to that, from what I can tell the sensitivity is not nearly as well behaved and predictable as shown. Same with the chips I have been playing with, that read wider bandwidths (e.g. XYZ)
So the idea is to provide a few reference illuminants to compare to chip readings and come up with a correction matrix or similar. So far I haven’t ventured that far, limiting myself to the two ends of the spectrum and interpolating. If one plays with WB sliders one realizes that accuracy in CCT is not critical, plus/minus a couple of hundred K is fine, though I think one could get pretty close with my setup. This is the CCT of my screen
And this is from a warm LED bulb, spec’d at 2700K: