pm-bench compare — diff two leaderboard JSON snapshots

[protosphinx]

Stacked on top of #15 (floor-baselines). Merge order: ... → #14 → #15 → this.

Summary

• New CLI verb pm-bench compare A.json B.json. Per-model score deltas as JSON; models unique to one side surfaced separately. Tasks/datasets must match.
• The "did my change move the numbers?" command. Useful both interactively (snapshot, edit, re-snapshot) and in CI (post the diff as a comment on PRs that touch scoring code).

What's new

• pm_bench/leaderboard.compare_boards(a, b) — pure function returning {"task", "dataset", "compared", "only_in_a", "only_in_b"}. ~25 lines.
• CLI: pm-bench compare A.json B.json emits JSON; exits non-zero with a clear message if the boards are on different (task, dataset) pairs.
• 6 new tests including a click-runner smoke and the cross-task rejection canary. 123 total.

Smoke

$ pm-bench compare \
    leaderboard/next-event/synthetic-toy.json \
    /tmp/snapshot-after-my-change.json | head -20
{
  "task": "next-event",
  "dataset": "synthetic-toy",
  "compared": [
    {
      "model": "markov-ref",
      "scores": {
        "n":    {"a": 158, "b": 158, "delta": 0},
        "top1": {"a": 0.9304, "b": 0.9504, "delta": 0.02},
        "top3": {"a": 1.0, "b": 1.0, "delta": 0.0}
      }
    },
    ...
  ],
  "only_in_a": [],
  "only_in_b": []
}

Test plan

• pytest -q — 123 passed (was 117 on PR Floor baselines: zero-time + empty-dfg — multi-entry on 3 boards #15)
• ruff check pm_bench tests — clean
• Identical-board diff produces all-zero deltas
• Tampered top1 surfaces as expected delta
• only_in_b surfaces newly-added entries
• Cross-task comparison rejected loudly

[protosphinx]

Merged into main as part of the audit-cleanup stack (commit 9c00b47). The full content of this PR is now on main.