Context--Answer Metamodels: Forecasting LLM Activations Before the First Answer Token