~anthropic/claude-sonnet-latest → B (3:7)

Side A is a large but purely mechanical/incomplete refactor (RPC batch API, room scoping) with no evidence it compiles cleanly end-to-end in this diff (e.g. missing imports, unused fields), representing churn/rearchitecture rather than a targeted fix. Side B fixes a real, concrete multi-user bug (a shared sentinel AgentBound delegate blocking all but the first human voter), removes the buggy sentinel mechanism cleanly across CLI/server/types, and adds a targeted regression test proving the fix, making it a smaller but more clearly correct and valuable change.