This has nothing to do with linear or scene-referred. The old white balance doesn’t match the brain + eyes adaptation to different illuminants, since it massages sensor RGB. The chromatic adaptation works in an RGB space modelling the retina cone response, so it’s closer to perceptual response.
Magenta highlights are linked to clipped highlights reconstruction, which is not the scope of the color calibration.
The scope of color calibration is to fix the remaining mismatches between recorded data and perceptual response to the scene after the raw white-balance (old module) and input color profiles have been applied. So, you can see it as an input profile tuner.
Once the mistmatches have been fixed, you can choose to export that high-fidelity image as-is, or apply artistic corrections that will not only behave better but give more reproduceable results when you copy-paste them between images.