The only exception to reverting to human control is if the software predicts there is a path to accident avoidance. In that case, the accident-avoiding action should be taken. That’s almost a tautology as it’s a given that FSD software will always _try_ to avoid accidents.
Culpability should be an entirely separate question. What are the expectations of use for a particular FSD software? Tesla is treading a fine line, but my understanding is that you are still expected to pay attention to the road and be ready to take control whenever using FSD.
Perhaps Tesla should also be modelling reaction times? If the time between reverting to human control and the moment of impact is too short, then you could argue the accident is a result of the software. However, what about the time between risk identification and impact time? Different people would identify a risk at different times. The same will be true for different self driving algorithms. FSD has a long tail problem. Very rare events are difficult to learn and test for. The best system we have for dealing with rare events is a human driver - so it’s more than reasonable to me to flip over control to drivers under these circumstances.