But it is consistent, right? Because if you sample infinite light direction you can get the one you want.

No, because the probability of hitting a point light is zero (no matter how many samples you take).

Oh~I forgot we are talking about the estimator's consistency! Thanks!