Add copilot calibration measurement

[JuneQQQ]

Summary

This cycle's competitive research → innovative increment.

Research. The WOLF benchmark (arXiv:2512.09187, Dec 2025) scores
werewolf-suspicion calibration with the Brier score. deepwolf is — in the
surveyed literature — the only werewolf project with an explainable
probabilistic copilot. So it is uniquely able to do something no competitor
can: report its own advisor's calibration back to the human who relies on it.

Shipped. A deepwolf calibrate command and deepwolf.copilot.calibration:

• plays many seeded games; at each daybreak, asks the copilot — from every
  surviving villager's seat — for its suspicions, and pairs each with ground
  truth (was that player really a werewolf);
• reports the Brier score (exact), the Brier skill score (vs a
  base-rate forecaster), the Murphy decomposition (reliability / resolution
  / uncertainty) and a reliability diagram.

Sample run (60 games, 7 players): Brier 0.206 vs baseline 0.226 — skill 0.087;
reliability 0.001 (the copilot is well-calibrated in the mid-range: it says 33%
and 34% turn out wolves, says 50% and 50% do) but slightly overconfident above
60%. The tool surfaces exactly that kind of finding.

Checklist

• ruff / mypy / pytest (81 tests, 14 new) all pass
• No behaviour change to existing code — purely additive
• CHANGELOG.md and README.md updated
• Research cited in the module docstring and the changelog

Notes for reviewers

Predictions are collected via the engine's observer hook at DAY_BREAKS — no
engine change needed. Only village-aligned viewpoints are scored (the copilot's
deductive job is a villager's job). The Murphy decomposition is the binned
approximation; the headline Brier score is computed exactly from the raw pairs.

[JuneQQQ]

Code review (post-merge record).

• Research-driven and properly scoped: WOLF uses the Brier score for agent suspicion calibration; this adapts it to deepwolf's distinctive asset — the explainable copilot — which is a genuine extension, not a copy. ✅
• Purely additive: no existing module is touched, predictions are harvested through the existing observer hook at DAY_BREAKS, so no game, seed or transcript changes. ✅
• The metrics are sound: Brier is computed exactly from raw pairs; baseline Brier equals base_rate*(1-base_rate) (verified algebraically — that is exactly a base-rate forecaster's Brier); the Murphy decomposition is correctly labelled as the binned approximation. The decomposition test asserts the reconstruction to a tight tolerance. ✅
• Only village-aligned viewpoints are scored — correct, since a werewolf trivially knows its packmates and would inflate calibration. ✅
• Empty-report path is handled and tested; skill_score guards division by zero. ✅
• CI green on 3.10-3.12, 81 tests.

The sample run is genuinely informative — it shows the copilot is well-calibrated mid-range but overconfident above 60%, which is a real, actionable finding. Merged via squash.