The QFN and BGA fanouts never read the net-class and item-type scoped .kicad_dru clearance rules that the route, diff-pair, plane and oracle steps all honour. A fanout can lay an HV escape next to another HV net at the layer clearance instead of the 0.6 mm the rule asks for, and later steps then hold the rule against copper that is already too close. Found by reading code, and nobody has measured it on a corpus board yet.