llm: headless batch CLI over runCellAsync (mock e2e pinned to runCell; JSONL out; .env keys)

llm: providers.js — anthropic/openai/ollama chat via bare fetch (UMD, injectable fetch)

llm: observe.js — coordinate-labelled board text + MOVE parsing (UMD, shared node/browser)

humanMove resolves temptations like the sim path (C10) — human Maintenance no longer stuck at 0

redesign: incentive-compatible metric + factorial RxGxE + peer rule-following MCTS + swap + fidelity fixes (Discovery/opponent-invariance/C*)