openai/gpt-chat-latest → A (4:1)

Side A makes a user-facing correctness fix by generating vote URLs from `display_path()` instead of stored URLs, keeping links consistent with the displayed DSL, and substantially strengthens coverage by exercising all 45 pairwise votes and asserting the final ranking through `GetGardenRank`. Side B mainly refactors recent-vote storage from a deque to an append-only list with query-time capping, removes trimming logic, and updates tests for that behavior, but the patch is largely an internal representation change without an equally clear project-wide functional improvement.