Activation Manifold Projection: Liberating Task-Specific Behaviors from LLM Architectures