openai/gpt-chat-latest → B (3:2)

Side B makes a broad architectural improvement by separating canonical path handling from identity parsing, introducing dedicated `canonical_path.rs` and `identity.rs`, removing identity rewriting from the reducer, enforcing strict stored-form identities without `@` at API boundaries, and updating APIs, HTML, types, and tests consistently. Side A meaningfully improves pair selection with new bridge and rank-refinement heuristics plus targeted tests, but its impact is confined to one algorithm, whereas Side B establishes cleaner long-term boundaries and data handling across the project.