An important point is that people have to feel that the punishment is connected to the consequences. If drunk driving was punished as harshly as killing when driving drunk, questions would arise such as: "does the life have no value?" or "if I am already driving drunk, does it make no difference whether I kill somebody?". People punished for killing in such cases would feel that they are punished for breaking the law, not for actual killing. Of course if people were perfect bayesians all these objections would fade away, but we are not.

This is connected to the testability. Our risk estimates are often badly wrong, and to base the justice on something regularly wrong seems not to be a good idea. When bad consequences are punished, one can end in jail just by having bad luck, but at least the biases and prejudices of the judges (in ideal case) play no role. If, on the other hand, risky behaviour was punished, then we would have a system possibly more consistent with theoretical consequentialist ethics, but far less objective.

By the way, the intuition which tells us that a wrongdoing has to be punished more severely if there are actual bad consequences is, I believe, an adaptation which prevents punishment based on biased risk estimates.