It sounds like you have a good grasp on one of the major differences:
Once you are manually picking a small subset of setups to check via simulation - there is no way to know if any of the setups are “good” or not. You can only know which of those simulates highest. Lots of times that is acceptable - but sometimes… it’s not!
On top of that, I would argue that comparing small numbers of data points with a simulator is not a good way to use simulation. I talked about it some in this post:
I would then argue that the solution that our optimizer gives is, at worst, just as good as anything you would get manually comparing specifically simulated sets. It is really hard to explain why set A simulating 1% better than set B doesn’t mean that set A is definitely better. That result more likely means: use either one, as they both fall within the bounds of the probable structural error in our simulation model.