Measure metric-convention choices on the corpus before committing

When picking set vs multiset vs ordered matching for tool-use precision/recall, do not argue it abstractly: compute all three over the real traces and count how many records actually separate them. Overlap differs only in the numerator (set intersection, Counter min-overlap, or LCS length) so the three are cheap to run side by side. A convention that never diverges from a simpler one on the data is unobservable there, which is itself the finding worth reporting; in one corpus 4.3 percent of traces had repeated tool calls (set vs multiset) but 0 percent had right-tools-wrong-order, making ordered and multiset numerically identical. Pair the corpus numbers with a small synthetic table (retry loop, reversed sequence, omission, extra call) since real data often contains no single record that separates all three.