perf(gfql): hop's full loop re-dedups the accumulated matches every hop (O(hops × matches)) on labeled / min_hops shapes #2047

Found in the #2036 dedup audit. graphistry/compute/hop.py full loop (the path taken when fast_path_enabled is off or the shape needs edge/node hop labels or min_hops) accumulates results with concat([matches_edges, hop_edges[[EDGE_ID]]]).drop_duplicates(subset=[EDGE_ID]) (hop.py ~632) and the same pattern for node labels (~685/695) on every hop, so each hop pays a hash pass over everything matched so far: O(H × M) for H hops. The fast loop already keeps visited sets (_domain_union / _domain_diff, hop.py ~566-576) and only touches the new frontier.

Not a wrong answer (contracts pinned by the hop semantics and rediscovery suites); a cost class distinct from #2036's removed O(E) dedup. Fix shape: move the full loop to the same visited-set accounting and pin it by extending test_hop_scaling_pin.py to a 4-hop labeled shape (cost must stay a bounded multiple of the fast loop). Owner priority: after the 0.60 stacks land.