All I can say is your problematic sample is very peculiar. The light source is suspicious in the first place, it’s around 2400 K with seemingly low color reproduction index. Since I don’t know anything about the conditions of the shot, it’s hard for me to conclude where the failure is.
For cases like this, the white balance module is more robust. As I said above, more robust usually means less accurate, but predictably inaccurate.
To solve this kind of issue, I guess we will have to perform chromatic adaptation in spectral (which is planned for the future, I have all the coding bits required, just need to wire them).