Quote: "What I don't understand is why they went with a left-hand system. I'd expect positive Z to come toward the "viewer.""
What value would then be at the camera position? Something exceptionally high? Remember that using IEEE754 floats the accuracy of the values isn't linear as with integers, but you have higher decimal precision at lower exponents as you have a finite amount of significant figures the float will be accurate to. So if the camera is at 0,0,10000000 then the Z-depth accuracy at that range would be abysmal.
I think the system makes perfect sense, once vertices of a model are transformed anything at 0,0,0 is directly on the camera, anything down the Z axis will be farther from the camera, X to the sides and Y vertically.