openai/gpt-5.3-chat → B (4:1)

Side A fixes several concrete test/mock issues (e.g., correct query splitting with regex, using getRequestBody instead of getInputStream, guarding null tokens, and preventing handler crashes with try/catch), restoring E2E auth flows. Side B introduces a substantial architectural improvement: a new async settlement worker with batching, cached ranking computation (ranked_items_cached), and removal of write-lock recomputation, which meaningfully improves performance and design beyond a simple fix.