~anthropic/claude-sonnet-latest → B (8:2)

Side B fixes an actual correctness bug in the core ranking algorithm (wrong divisor causing bipartite oscillation and incorrect rankings on star topologies), grounds the fix in a cited paper, and adds thorough regression tests (Rust unit test plus Clojure fixtures covering star/inverse-star/chain/cycle topologies). Side A is mostly a demo-counter removal/refactor plus introduction of an async settlement worker with caching, which is reasonable housekeeping but doesn't fix a bug and carries more churn/deleted test coverage without a comparably rigorous justification.