~anthropic/claude-sonnet-latest → B (3:2)

B fixes a genuine, well-diagnosed algorithmic bug (bipartite Markov chain oscillation in rank-centrality scoring) with a principled correction grounded in cited literature, plus targeted regression tests across multiple topologies. A adds a large, plausible-but-sprawling invite-system feature with meaningful surface area (server state, RPC, CLI, reducer, timeline) but is riskier/less proven, includes speculative unused fields (InviteMinted/InviteRedeemed events never emitted, room_timeline/ActiveInviteState seemingly unused by the new timeline.rs), and its correctness is less rigorously validated than B's mathematically justified fix.