Mitigating Safety Tax via Distribution-Grounded Refinement in Large Reasoning Models Paper โข 2602.02136 โข Published Feb 2 โข 7