Add a game simulator for generating synthetic games

[jeremymanning]

Deferred from the initial scoping of #1, which is analysis-only. Filing so we don't lose it.

Idea

Implement a D12-rules game engine that can play games start to finish with configurable agents,
producing logs in the same schema the scraper emits. Since the rules are the same whether a human
or a bot plays, board-dynamics findings from simulated games transfer to real ones.

Why it's worth doing

• Null models. Most of the interesting claims in basic approach #1 are comparative: is this map unbalanced,
  is this position-strength trajectory unusual, do these games cluster more than chance? Each of
  those needs a generative null to compare against, and a simulator is the principled way to get
  one.
• Counterfactuals. Replay a real game from turn k with a different move and measure the
  swing in position strength. Turns descriptive metrics into causal ones.
• Metric validation. A position-strength metric should predict wins in simulation. If it
  doesn't, it's wrong, and we'd rather learn that from cheap simulated games than from the paper.
• Coverage. Real data will be thin for rare configurations (unusual maps, player counts,
  drop-out patterns). Simulation fills gaps.

Scope sketch

• Rules engine: territory placement, reinforcement/bonus calculation, attack resolution, fortify,
  card sets, elimination, turn order
• Agent interface with a few baselines (random-legal, greedy-expansion, bonus-completion,
  border-minimizing), plus a hook for learned agents
• Emits logs conforming to the same schema as scraped games, so the whole analysis stack runs on
  simulated and real games interchangeably
• Validation: agent strength ordering is stable, games terminate, distributions are sane

Explicitly out of scope here

Negotiation between agents. LLM agents talking to each other is a different project answering a
different question — see the companion issue on external conversation corpora. Board play only.

Depends on

#1 — needs the log schema, map/graph representation, and position-strength metrics to exist first,
since the whole point is to plug into them.