hiyenwong/ai_collection/arxiv-2608-23300v1-evaluating-sat-solver-metrics-as-predictors-of-humskill versions