Context windows: 1M tokens vs usable memory

I've been testing long-context models on real codebases instead of needle-in-a-haystack benchmarks. Feeding 400k tokens of a monorepo