review-branch skill validation fails for unfetched remote refs in worktrees

[cderv]

The review-branch skill's input validation (step 1) runs git rev-parse --verify -- <branch> to check that the --base ref is valid before proceeding. This fails when the ref is a remote tracking branch that hasn't been fetched locally yet — common in worktrees where remote refs may not be present.

Observed twice when running from a Claude Code worktree:

$ git rev-parse --verify -- upstream/main
fatal: Needed a single revision

The skill spec says to abort on validation failure, but in practice Claude improvised a fetch + retry:

$ git fetch upstream main
$ git rev-parse --verify upstream/main
aef54133cf0b3f621243cbd5d7ad90e642cad819

This recovery isn't guaranteed — it depends on the model choosing to retry rather than following the "do not proceed" instruction.

The validation step could handle remote refs (matching <remote>/<branch>) by fetching before validating, or by skipping the strict check for refs that look like remote tracking branches and letting roborev review --base handle validation itself.