Actionable Activation Directions for Detecting and Mitigating Emergent Misalignment Across Language Model Families