Where Rounding Happens Matters: Rounding-Location Alignment Stabilizes LLM Reinforcement Learning