Orthogonal activation steering frameworks and weights freed from behavioral alignment filters, covering both hybrid instruct and reasoning models