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

B fixes a real security/design flaw (feed cutoff keyed by unauthenticated actor string) with an auth-bound delegate model, adds thorough integration tests covering rejection paths, and updates CLI/docs consistently. A is a substantial refactor adding children-fetch support and SSE-as-JS-eval, which is useful but more speculative/UI-plumbing with less clear-cut correctness payoff and more churn (rewriting SSE protocol, JS handling) than B's targeted security fix.