Inducing self-other overlap with SFT reduces deception at scale, but generaliza