Ablating individual neurons within attribution circuits can identify whether specific input tokens cause a persona steering vector to fail