openai/gpt-chat-latest → A (4:1)

Side A fixes a production-facing reliability issue by requiring OAuth when credentials exist, retrying once with a refreshed token on 401/403, propagating token acquisition errors instead of silently falling back, and improving error handling with truncated diagnostics and a regression test. Side B corrects an ordering bug by moving the zero-ratio early return before `ensure_item` and `voted_pairs.insert`, preventing ghost state, but it is a smaller, localized fix with narrower impact.