The paper's headline retrieval comparisons obey four explicit invariants stated in the Strict-Parity Evaluation Contract: (i) an identical query encoder and identical dense seed pool size $m = 100$ for all dense-seeded systems; (ii) an identical retrieval cutoff $k = 10$; (iii) identical corpus and split; and (iv) an identical matching rule, specifically exact concept-ID comparison of each retrieved item against the gold concept IDs. Together these invariants instantiate the strict-parity (SP) contract at the level of concrete numbers, so any difference between two compared systems in the headline numbers must come from the graph component itself rather than from an interface mismatch.