Calibrate evidence quality gate thresholds

Add bootstrap confidence intervals to evaluation metrics

Fix EDA review-level rating distribution and resource cleanup

Fix eval workflow and align README metrics with eval_results JSON