~anthropic/claude-sonnet-latest → B (6:4)

B builds substantial lasting architecture (a fractal ItemId/GlobalTree model, breadcrumb navigation, parent-child ranking, journal worker replacing settlement) that generalizes the app beyond flat subreddit scopes, while A merely replaces an over-engineered but self-contained autocomplete graph with a simpler paste-and-go parser. However, B's diff also carries real risk/incompleteness (renamed settlement->journal with duplicated logic, legacy scope shims, less test coverage of new tree semantics), so it's not a clean sweep, but its net design contribution outweighs A's more isolated simplification and deletion of test infrastructure.