LLM Werewolf runner adds skill memory

This is more interesting as an agent-behavior sandbox than as a game demo. The repo surfaces a real weakness in LLMs: they can reason through a swap, then still emotionally anchor to the wrong identity.

• –The persistent `gameskill` file is a lightweight memory loop, so each match can influence the next without finetuning.
• –Making the runner API-agnostic lowers the barrier to testing older local models that do not support tool calls well.
• –The live transcript plus phase machine gives you a clean trace of how identity confusion, bluffing, and strategy evolve over time.
• –The project is useful for studying prompt design and state tracking, especially because “goal-oriented” prompting already changed model behavior.
• –A 5-player version would probably increase entropy and make the social deduction dynamics less deterministic.