But isn't the intention of the synthetic training to make sure that the edge of the envelope always remains out of reach? That actions are taken so that stalls and upsets just don't happen?
If they do, then a simulator certainly can't wander beyond the edge of the gathered test flight data to reproduce them. If a large part of the aircraft falls off, or it enters a stall, then the simulator won't be able to reproduce the experience unless the model contains data as extreme as that event.
Aerobatic training might be very interesting, but would it have helped AF447 or Colgan whatever in Buffalo? I don't really think so. Wasn't it some exotic maneuvering training at American that could have contributed to the actions leading to the Queens NY Airbus crash in 2001?