Add bidding-based discussion turn-taking

[JuneQQQ]

Summary

Implements #19 — bidding-based discussion turn-taking, the highest-value item
from this cycle's competitive research.

Research that motivated it

Google's Werewolf Arena (arXiv:2407.13943) introduced a dynamic turn order
where agents bid to speak instead of following a fixed seating order — it
better mirrors real discussion. deepwolf used fixed seating order; every serious
comparable project has moved past that.

What this adds

• GameConfig.discussion_mode — validated, "ordered" (default) or "bidding";
  a --bidding flag on simulate and play.
• engine._speaking_order() collects a bid from each living agent, emits each
  as a public SPEAK_BID event, and seats speakers highest-bid-first (ties by
  seat).
• Agent.bid() — new concrete method; RandomAgent, LLMAgent and the CLI
  HumanAgent override it. Bilingual prompts; the offline mock handles the new
  bid decision kind.

deepwolf's innovation over the original

Werewolf Arena's bids are an internal mechanism. Here bids are public — a
priority and a reason, emitted as events. An eager bid backed by a thin reason
is itself a tell that other agents and the copilot can read. That fits
deepwolf's "everything is explainable" principle.

Checklist

• ruff / mypy / pytest (67 tests, 7 new) all pass
• Default is "ordered" — existing games and English output unchanged
• CHANGELOG.md and README.md updated

Closes #19

[JuneQQQ]

Code review (post-merge record).

• Research-driven and properly attributed: the PR cites Werewolf Arena and adds a genuine deepwolf-specific twist (public bids) rather than copying — exactly the brief. ✅
• discussion_mode defaults to "ordered"; verified the ordered path emits zero SPEAK_BID events and English output is unchanged, so no existing game shifts. ✅
• _speaking_order clamps priority to 0-10 and truncates the reason to 200 chars — a rogue/hallucinating bidder can't distort or flood the round. A test covers it. ✅
• Agent.bid is concrete, so third-party agents need no change; the mock handles the new bid kind so offline games still run. ✅
• Bids carry both priority and reason in event.data, so the copilot can consume them later (good seam for the follow-up). ✅
• CI green on 3.10–3.12, 67 tests.

Clean increment. Merged via squash.