test(eval): publish a reproducible Terminal-Bench score

[QiaolongLi1201]

Goal

Replace readiness claims with a reproducible public Moss Terminal-Bench result when the official submission path is available.

Scope

• Pin the benchmark revision, task manifest, Moss adapter revision, model, reasoning settings, and container image digest.
• Run the existing 200-case harness preflight before any submission.
• Record environment/tool versions, pass/fail/skip totals, cost and latency telemetry, and raw artifact checksums.
• Keep credentials in environment/configured runners; never commit them or embed them in artifacts.
• Publish the exact reproduction command and link the official leaderboard/submission result.

Acceptance criteria

• A clean checkout can reproduce the submitted run from documented commands.
• The official score, raw results, and Moss commit SHA agree.
• Skips, retries, and infrastructure failures are reported separately from task failures.
• The README/user guide makes no superiority claim beyond the measured result.