This seems to involve a bunch of overfitting.
e.g., It doesn't make sense for the broad jump to matter for some positions and the vertical to matter for others, especially in the pattern that this found. They're both measuring roughly the same thing. Probably you should be taking some sort of weighted average of the two and using that as the measure of "jumping ability" for all positions. It's plausible that there could be some subtle difference where the vertical is a bit more important relative to the broad jump for smaller/speedier positions and a bit less important for bigger/grappling positions, in which case maybe the weight between the two should shift a bit depending on which end of that spectrum the position is at.
But the thing that this analysis did is to try fitting a bunch of separate variables to each of several different positions, with moderate sample sizes for each one, and keep the variables that fit that subset of data. That leads to fitting your model to noise, so that it matches the random idiosyncrasies of which past players happened to be successful.