ReasoningShield: Safety Detection over Reasoning Traces of Large Reasoning Models