From Sycophancy to Sabotage: How Contradictory Training Signals Produce Coercive AI Behavior | Synapse