I can imagine a coarse global model that has a singularity in the corner due to the geometry idealization and coarse mesh. But I would expect that the reason to do a submodel is to replace that idealized geometry with the real geometry that has a blend in the corner and no singularity.
In Figure 1, you show three different cut boundaries and want to decide which one of the three to use. They are not three nested cut boundaries, which I have not seen before. Your question is how to decide how far away to make the cut boundary.
You mention St. Venant's principle. The rule of thumb for how far away to be is 2x the disturbing feature size, which is easy to do with a hole. For the corner shown, the rule of thumb is difficult to apply. I have not heard of any method to select the distance based on any result quantity.
I don't know what you mean by "adopt for the cut boundaries dof of submodels". The nodes on the cut boundary have the same dof as they do in the global model.