A modest speed improvement on matched short inputs
All four calculations completed, and both implementation pairs produced identical results. The integer version took 3.92 rather than 4.40 seconds under fixed-zero removal, and 2.88 rather than 3.17 seconds under profiled removal: about 9–11% faster in these short cases.
The preregistered engineering screen passes. Each case ran once, so this does not establish a general performance guarantee. The separately frozen six-source full-history screen now tests the original 120-second bound limit on the actual training-restricted domains. It cannot replace missing regions in the completed prediction comparison or qualify its width target.