Surface LLM agent reasoning into the event log

[JuneQQQ]

Summary

Competitive research (this cycle) turned up the WOLF benchmark (arXiv:2512.09187), which makes per-agent chain-of-thought scratchpads central to explainable LLM social-deduction evaluation. deepwolf's LLMAgent already collected a reasoning_log per decision — but it was never exposed. Saved transcripts recorded what each agent did, never why.

This PR closes that gap.

What it does

• New EventType.AGENT_REASONING, private to the deciding agent (visible_to = {actor}). Other agents' views are byte-for-byte unchanged.
• Agent.last_reasoning() is a concrete default returning None; LLMAgent overrides it. Baseline agents (Random, etc.) emit nothing — no noise for non-LLM games.
• engine._emit_reasoning() is called after each choice action: kill, inspect, protect, vote, shoot. Empty / whitespace reasoning is dropped.
• Bilingual: a new agent_reasoning i18n key (en + zh). Chinese games render the reasoning prefix in Chinese.

deepwolf's twist over WOLF

WOLF keeps reasoning in out-of-band scratchpads. deepwolf threads it into the event log itself, so the reasoning is rendered, replayed and round-tripped through the same machinery as every other event — including the JSON transcript and the upcoming deepwolf replay (#23). The visibility model means an agent's reasoning never leaks into another agent's view.

Checklist

• ruff / mypy / pytest (89 tests, 6 new) all pass
• No change to existing event types — purely additive
• English event text unchanged (no existing message touched)
• CHANGELOG and README updated

[JuneQQQ]

Code review (post-merge record).

• Visibility is correct: AGENT_REASONING is restricted to visible_to={actor}, and verified by a dedicated test. Other agents' views (and therefore the copilot's belief model) are unchanged. ✅
• Concrete-method opt-in: Agent.last_reasoning() returns None by default, so baseline agents emit no events — no noise for RandomAgent-only games. Confirmed by test_random_agent_emits_no_reasoning_events. ✅
• Whitespace guard works at both layers (LLMAgent strips internally, engine strips defensively). Test test_empty_reasoning_string_is_not_emitted exercises the engine's guard. ✅
• Bilingual: new agent_reasoning i18n key; Chinese smoke run shows "推理（kill）：..." correctly. The decision-kind label (kill/vote/...) is left as a structural string — matches how phase is stored. ✅
• The transcript round-trip via to_json carries reasoning end-to-end, so deepwolf replay (feat: add deepwolf replay command #23) will naturally surface it once that PR lands. Nice composition with another open PR. ✅
• CI green on 3.10–3.12, 89 tests.

Merged via squash.