Within the strict-parity contract, the paper distinguishes two cases that are not part of the headline comparison. Systems that add an extra learned reranker on top of the shared first-stage are still reported, but their numbers are interpreted separately from the headline strict-parity numbers under the SP+ extension because the extra learned component changes the retrieval interface. Best-effort baselines that violate the contract (different encoder, different $m$ or $k$, different matching rule, or different split) are moved to the supplement rather than placed in the headline table. This reporting policy is what keeps the headline numbers attributable to graph policy alone.