~anthropic/claude-sonnet-latest → B (6:4)

B fixes a real correctness bug (ghost items and falsely-marked voted pairs from zero-ratio votes) with a minimal, precise change and updates the test to verify the actual fixed behavior. A adds substantial heuristic complexity to pair selection with speculative tie-break priorities and several new tests, but it's tuning/heuristic refinement rather than fixing a concrete defect, and its correctness/value is harder to verify as clearly beneficial.