openai/gpt-chat-latest → A (4:3)

Side A changes the sibling navigation behavior so each unranked sibling becomes its own navigation group instead of all unranked siblings being merged together, matching the documented ranking model, and adds a targeted regression test to lock in that behavior. Side B substantially improves the local development fixture workflow with persistent data, cargo-watch integration, preferred port selection, and summary reuse, but those changes primarily affect developer iteration rather than core product behavior.