Hi Jonathan,
as Wishgranter said, there is no single answer. It depends on the camera and the object and the images.
So you need to try it out. I used the last in the menu quite a bit and got good results. Mine is a small, cheap camera, so that confirms what WG said.
Just use a small fraction of your images of one spot (say 50) and make comparisons.
And yes, RC DEFINITELY uses information from previous alignment, IF they are present as components.
But I also am convinced that there is a randomness in the algorythm somewhere, so each alignment will lead to slightly different results. I am talking about internal mechanisms though, not necessarily noticeable differences. Altough even that I have observed.
I think that the technology is not something that has one right method, but it needs experience to get a feeling for when the results are as good as they are going to get.