That's an advantage over linear models, but GBTs handle non linearly-separated data just fine. Each individual tree can represent an arbitrary piecewise-constan

That's an advantage over linear models, but GBTs handle non linearly-separated data just fine. Each individual tree can represent an arbitrary piecewise-constant function given enough depth, and then each tree in turn tries to minimize the loss on the residual of the previous trees. As such, they're effectively like a neural network with two hidden layers in terms of expressiveness.