The point that a big enough window would still fail is the one most introductions skip, and it is what makes RAG an architecture rather than a workaround.

Attention over a long context is not uniform, so the retrieval decision does not disappear when the window grows. It moves from an explicit ranked step you can measure into an implicit one inside the model that you cannot. That is the trade I would put in front of anyone arguing long context makes retrieval obsolete, and it is about observability rather than cost: with a retriever you can ask which chunk was fetched and at what rank, while with everything stuffed in the prompt you get the answer and a guess.