G x E

It is perhaps time to try alternative models to the one they used. Moderating the regression may be reasonable when the source only affects one trait, because the factor variance and the regression are confounded, so the moderated regression is a proxy for change in the latent factor variance (heteroscedasticity). However, mechanistically, one might expect either the variance of the latent variable to be increasing, or the sensitivity of the target variable to be increasing. If the latent common factor variance (A C or E) changes then both the target trait and the moderator should show variance increase. If the sensitivity of the target variable is changing then both the common factor and the specific components should show increased regressions, and perhaps by the same amount.

LOL re: the tweet you received. The paper doesn't show this, it says "good genes will out" - once the teacher quality is high, what is left is individual differences in cognitive ability, which are largely genetic in origin. When the teacher quality is low, genetic factors don't get so much of a say in the matter. Possibly, variation in the domestic environment has more impact, although there's not much sign of C or E increasing at the low end of the distribution of teacher quality.