When More Data Hurts: A Directional Goodhart Condition for Multi-Stakeholder Preference Learning | Synapse