Evaluate and migrate carve-changesets remediation to review-fix-loop

[shaug]

Outcome

Evaluate and migrate the review/remediation portions of carve-changesets to the packaged review-fix-loop skill while preserving changeset composition and publication semantics.

Design: design/review-fix-loop.md
Related epic: #95

This is an intentionally separate fast follow, not a subissue of #95.

Scope

• Identify the local and published review/fix paths that should delegate.
• Reuse the proven babysit-pr integration for published changesets where appropriate.
• Preserve ordered changeset ownership, reference-branch safety, and final-tree equivalence.
• Document any path that should remain caller-owned rather than forcing delegation.

Non-goals

• Reworking the changeset carving algorithm.
• Making review-fix-loop aware of changeset stacks.

Acceptance criteria

• Each current remediation path has an explicit delegate-or-retain decision with evidence.
• Delegated paths map all terminal states safely.
• Published paths do not bypass the established PR watcher integration.
• No intermediate fix is pushed before convergence.
• Existing final-tree equivalence and ordered-stack evaluations continue to pass.

Dependencies

Blocked by the babysit-pr migration, which is itself blocked by the packaged-skill leaf of #95.