openai/gpt-chat-latest → A (4:1)

Side A implements a substantial architectural refactor: it replaces cached `GroupState` with a simpler `ScopeVotes` model, derives ranking edges and connected components on demand, updates persistence and ranking algorithms, and adapts callers and tests throughout the project. Side B adds a sizable URL canonicalization/graph subsystem with extensive tests, but it is largely an isolated new feature, whereas Side A simplifies core state management and removes redundant cached data in a way that affects the project's central ranking pipeline.