Assessing the overall and partial causal well-specification of nonlinear additive noise models