Add per-run score fingerprinting for contest reproduction

[jaisinha77777]

Summary

Addresses #271.

`run.py` writes a full structured event log per run, but the actual per-agent `focal_scores`/`background_scores` dict is only printed to stdout and then discarded — there's no durable, parseable record of exactly what a completed run scored. That makes the concern raised in #271 concrete: two runs can land on the same aggregate Elo while actually differing per-scenario per-agent, and there's currently no way to tell those cases apart after the fact.

Changes

• `run.py` now also writes `simulation_scores_{game}_{scenario}.json` alongside the existing HTML/JSON structured log, containing the per-agent focal/background scores plus run identity (game, scenario, model name, api type, UTC timestamp, and the exact command invoked).
• New `fingerprint_run.py`, a companion script that reads one or more of those score files and computes:
  • `content_hash`: a SHA-256 Merkle root over every `(scenario, role, agent, score)` tuple across the supplied runs. This is what makes drift detectable at a finer grain than aggregate Elo — two runs with the same total score but a different per-agent split, or the same scores under a different scenario, produce different `content_hash` values.
  • `bundle_hash`: `content_hash` combined with each run's identity (model, timestamp, command), so two runs with identical scores but different provenance are distinguishable from runs that are genuinely identical.

Usage

python -m examples.games.run --game=haggling --scenario=fruitville ...
# writes /tmp/simulation_scores_haggling_fruitville.json

python -m examples.games.fingerprint_run \
    --scores /tmp/simulation_scores_*.json \
    --output /tmp/fingerprint.json

Test plan

• `examples/games/fingerprint_run_test.py`: 19 tests covering Merkle-root determinism/order-independence, sensitivity to per-agent score changes and to scenario changes, the "matching aggregate but different per-agent split" case Contest run provenance: making model drift detectable with per-run output fingerprinting #271 specifically calls out, focal/background collision-avoidance, bundle vs content hash independence from provenance, glob/path resolution, and the CLI end to end (via `main()` with a temp dir, no simulation run required).
• `python -m pytest examples/games/fingerprint_run_test.py -q` → 19 passed.
• Manually exercised the CLI end-to-end against a synthetic score file (no LLM/API keys required) and confirmed the output JSON report.
• `python -m pyink --check` on all changed/new files → clean.

Open questions from the issue

The issue itself left a few questions for maintainers unanswered (whether the original 2024 contest logs exist to compute a reference hash against, and whether this belongs in #265 instead). I didn't block on those — this PR is scoped to the mechanism itself (computing and persisting the fingerprint), which is useful regardless of how those get resolved. Happy to adjust the file layout/integration point based on maintainer feedback.