This has been an interesting discussion. So..... I did some thinking... and putzing with a speadsheet "model."
As far as I can ascertain, SGP is the only focus software that uses the whole field for focusing. My conclusion is that if the stars are equally distributed, the focus algorithm would indeed converge like the one on the right (above with minimum mean). I was partially wrong.
However, if the star sample had mostly stars around the edge of the screen and a deep sky object in the middle, the focus in the center would be off.
I've watched the behavior of the auto focus. If you have a decently long exposure, the star sample will be distributed evenly.... usually.
"Forcing" the sample to the middle would reduce the probability of the focus algorithm homing in on stars that are skewed in only one "bad" area (like the edge)
Still, the subframe feature would allow some real life experimentation and also help with the globular issue. Two things for the price of one.
This is not a trivial thing to understand. Hurts my head sometimes.