Lost-in-the-Middle Is Still Real in 2026 (Even on 1M-Token Models)

Long-context models still drop facts buried in the middle of their input. Here is a 40-line eval that proves it, and three fixes that work.