Rank agent versions against the fewest judge verdicts

An LLM judge compares agent versions in pairs on the same tasks and picks a winner each time. Its verdicts are not always consistent: A can beat B, B beat C, and C beat A. The release review still needs one ranking.

Write rank_versions(versions, verdicts) that returns the versions in ranked order, best first.

• versions: the version names, at most 10, all different. Names use only letters, digits, . and -.
• verdicts: a list of [winner, loser] pairs, one per judge comparison. The same pair can appear many times, in either order.
• Return a list that holds every version once.

Rules:

1. A ranking goes against a verdict when it puts the loser above the winner.
2. Return a ranking that goes against as few verdicts as possible.
3. If several rankings tie, return the first one. Compare rankings name by name from the top, in plain string order: by character code, so B sorts before a.
4. Ignore a verdict that names a version not in versions, or names the same version twice.