xdotool chat injection silently fails under KDE Wayland/Xwayland, so the
v1 harness recorded runs whose scene commands never executed. v2 drives
scenes from a generated world datapack, detects crashes, aligns MangoHud
CSVs by file birth time, flags vsync/backup/low-n windows, and records
full run metadata to JSONL; analyze.py applies the BENCH-MATRIX win rule.
