Now consider another hypothetical where people smoke cigarettes due to their biases, and other people without those biases have a significantly higher incidence of being run over by buses. Then, the biases that cause smoking cigarettes are not "significantly harmful on average" as compared to the alternative.

If that were true, then the cognitive defect would be the inability to distinguish between the problems of choosing whether to smoke cigarettes and how to avoid being run over by buses. Both the "biased" and "unbiased" people are somehow throwing away enough information about at least one of these problems to make them seem isomorphic in such a way that the effective strategy in one corresponds to the ineffective strategy in the other. The underlying problem behind the bias is throwing away the information.