Gravar-mail: On convex least squares estimation when the truth is linear