AGENTS.md — benchmarks/ Benchmark scripts and performance evaluation. Keep runs reproducible and inputs explicit. Separate harness changes from algorithm changes. Preserve output contracts used by downstream analysis.