R-squared inference under non-normal error