In my opinion, once you went through the trouble of estimating the per-particle scale, the scaling should be preserved. In most cases, I would think that the scale should be estimated for the highest-resolution consensus reconstruction, and then be kept after classification into states.
The only scenarios when you would want to reset the per-particle scales that I can think of are for low-resolution reconstructions where estimates might be off (but then estimating per-particle scale does not make sense in the first place) of if you have wildly different states/complexes and the minor states will be artificially suppressed in the consensus reconstruction.