openai/gpt-chat-latest → A (3:2)

Side A substantially refines the pair-selection algorithm by introducing structured component layout tracking, prioritizing attachment of unranked items to established components, and adding rank-aware 'zip' refinement once the pool is connected, with multiple new tests covering the behavior. Side B fixes a real correctness bug by moving the zero-ratio early return before item registration and voted-pair insertion, preventing ghost state, but it is a narrowly scoped fix compared with A's broader, lasting improvement to core ranking behavior.