Adaptive and Explicit safe: Triggering Latent Safety Awareness in Large Reasoning Models