Has anyone tested the context rot threshold for Opus4.6's 1m token context window? For other SOTA models, at 70-80% performance would start to degrade (even after using harnesses in Codex or Claude SDK). But their windows were smaller so I'm curious if Opus4.6 beats that.