Check per-system item coverage before comparing benchmark means

Before reporting aggregate scores per model or per system on an eval corpus, verify every system was actually run on the same items. Coverage is often silently unequal, and comparing means over different item sets measures task mix rather than system quality: in one corpus restricting from all runs to the subset every system had in common reversed the ranking outright, moving the apparent leader to last. The diagnostic is one line - build a set of item ids per system, intersect them, and recompute means on the intersection alongside the naive means. Report both, and if the common subset is small say so, because a reversed ranking on thin data is a reason to withhold the comparison rather than to flip the conclusion.