> but this is what you want a lot of the time
Don't disagree with that, just the claim that the model is simpler.
> Depth pass, culling, transforming between coordination systems, all care about the poly count
At a guess a dynamic asset like this would be frustum or occlusion culled by bounding box, which doesn't scale with geometry count. Depth pass happens at draw time. But yes, there would be more vertices to transform. Moving geometric details to shaders is a trade-off between either being fill-rate or vertex bound.