Why it is a question
THEORY-1158 reports Table 1 of LIT-1820 (read in NOTE-1331): composition through third vertices gains 26–45 points over a single path, and Wilson-loop routing adds 0.6–7.8 more. The released code was not run, so the accuracies are the paper’s own. The second half of THEORY-1158, the AUC ceiling set by exactly flat clean data, was recomputed by enumerating S4 and does not depend on a rerun.
What would answer it
A rerun of the released harness, or an independent implementation, that reproduces Table 1’s single-path, uniform and holonomy columns within their stated intervals, including the paired holonomy − uniform gains. A rerun on new synthetic data built flat does not bear on the arithmetic half of THEORY-1158, which is already checked.