Hi Pavel, I've identified the root issue of my original problem - I am using a reference model and the problematic side chains don't agree with the reference. Easy enough to fix.
Also I was using structure factors calculated by Buster, not realizing that this file contained HL coefficients based on my model. Phenix read these coefficients and switched to MLHL; the resulting instability is probably what led to the increase in Rfree. Re-refinements using the ML target show that the Rfree is essentially the same with and without RSR, to be expected since the structure is about finished.