openai/gpt-chat-latest → B (4:1)

Side B adds a new end-to-end browser test that seeds a pool, exercises the vote→next-pair workflow, and verifies both edge-history updates and that every displayed pair stays within the requested pool, providing lasting regression coverage for an important user flow. Side A mainly removes now-unused parameters and stops updating the vote-compare preview, simplifying the implementation but largely acting as cleanup with comparatively limited functional impact.