Multi-model review remains mandatory while the original bead, its acceptance criteria, and required guardrails now bound PR scope.
Fixed
- Every review root cause receives one finite disposition: fix now, block, reject with evidence, or one qualified follow-up.
- A model severity label alone never creates a Beads task. Follow-ups must be reproducible, actionable, non-duplicate, worth scheduling, and outside the current PR scope.
- Required defects remain merge-blocking regardless of severity.
- Review stops on a verified block or after at most three rounds, preventing recursive review-task churn.
- Canonical and generated skills, templates, policy mirrors, and regression coverage now share the same rule.
Full details are in CHANGELOG.md.