Why Generative and Discriminative Can Share a Boundary

A production system frequently has one sensor missing at test time, but your model family can factor the joint feature likelihood cleanly. Which side gets a practical advantage?