From hard refusals to safe-completions: toward output-centric safety training