refinement iteration - discrepancy from paper and potential bug

[sabahsima]

I'm looking into the code and there are 2 points in the implementation I wanted to confirm are as expected:

1. According to the paper, when a gaussian is being cloned, the cloned gaussian will be moved in the direction of the positional gradient. From what I can see from the implementation it is cloned in the same location, and the only difference is the exp moving average that is reset to zero in the cloned gaussian (which will change the optimization result, which is the critical part here). Is this the desired behavior? why is this implemented differently than what's written in the paper?
2. In the densify_and_prune function, the max_radii2D is being reset as part of the densification (cloning phase). This happens before the pruning check whether this radius is larger than a threshold, so this line is obsolete big_points_vs = self.max_radii2D > max_screen_size, which seems like a bug. This line does appear again in the prune function that is called in the end of the optimization, so the final gaussians will not include large gaussians, but according to the paper this should have been done in all refinement iterations.

Am I missing something?