Evaluate Juror in Martian's 50-PR review benchmark

[cderinbogaz]

Parent: #39

Context

The open Martian code-review benchmark evaluates major reviewers on 50 PRs and documents how to add another tool. An external corpus would provide stronger third-party comparability than Juror's own benchmark alone.

Reference: https://github.com/withmartian/code-review-benchmark/blob/main/offline/README.md

Acceptance criteria

• Reproduce the benchmark locally with a version-pinned Juror configuration.
• Record raw Juror outputs, model versions, preset, cost, latency, and run failures.
• Document any adaptation required to fit the benchmark protocol.
• Publish reproducible results without changing the benchmark's evaluation rules.
• Submit Juror upstream if the maintainers accept new reviewers.
• Link the external result from Juror's benchmarking documentation, including limitations and date.