Although I'm not a very regular user of modeller, I believe it is well
known that long gaps, especially terminal gaps, look strange when
modeled using default options.
Sure. Typically, we just remove those regions and don't model them. Any
modeling of these regions is just going to be fantasy anyway.
"if distance of residue x from center is greater than some huge value
(2-3x the predicted Rg, based on sequence length), add a bias to the
potential for x toward the center of the protein"
The easiest way to do that would be to add an upper bound distance
restraint between each C alpha and the center of gravity of the protein.
An example is at