The third step promised  in DM-18630 wasn't completed on that ticket. Opening a new ticket because I think this analysis would be useful:

Run2.1i uses a simple distortion model and a perfect reference catalog, which results in minimal improvement from jointcal because all information is being used in the single frame processing step. Rerun single frame driver + jointcal with an increasing sparse reference catalog to see how singleframedriver and jointcal degrade in relation to each other.