I vaguely recall a similar discussion from maybe a year or more ago. I think the gist was that registration can make subtle distortion corrections which can lead to some noise due to pixel interpolation. When stacking multiple independent sets that interpolation can be repeated and may not match up with other stacked sets. I think the recommendation was that you should use one common Reference Frame for all registrations of multiple nights and / or multiple filters.
I normally do something like that anyway. I take a good frame from the set of multiple nights and plate solve it. I then use that reference to generate a synthetic star field that is 15% (or so) larger than that initial frame. I register all frames from all sets to that synthetic reference frame. In theory, only a single set of distortion corrections are made that adhere to the one reference frame.
After integration, I crop the "ragged" edges and blank unused area of the over-sized integration. That is the starting point for further processing. I liken it to using a mosaic methodology even though I have never done a mosaic.