I’m not an expert, but if I’m correct about face normals and collosions, then that is not DIRECTLY related to “performance.” However, if deltaTime increases from a delay in processing a step (aka "performance), even for one step, then the impulse applied by the physics engine from collisions increases the delta velocity and therefore the predicted delta position beyond the collided’s faces.
So in summary, maybe tangentially related to performance. Under this theory, restricting step size seems like it restricts the size of the calculated delta position and recognizes a collision.
Edit: @Cedric’s comment here is relevant