That it did well on held-back data should convince you that you don't understand overfitting.

They didn't do well on the gene level: Analyses of individual SNPs and genes did not result in any replicable genome-wide significant association

No, they didn't try to measure non-linear effects. Nor did they try to measure environment. That is all irrelevant to measuring linear effects, which was the main thing I wanted to convey.

No, the fact that you can calculate a linear model that predicts h_2 in a way that fits 0.4 or 0.5 of the variance doesn't mean that the underlying reality is structured in a way that gene's have linear effects.

To make a causal statement that genes work in a linear way the summarize statistic of is not enough.