这次修复关闭了聊天链路中的 Eino checkpoint 自动恢复，避免旧断点抢先恢复导致模型反复回复历史对话。

This update disables automatic Eino checkpoint resume in chat flows, preventing stale checkpoints from taking over and causing repeated old responses.

同时新增/完善了按轮次展示 Token 用量的统计能力，可查看每次对话的输入、输出、缓存、推理和模型调用次数。

It also adds/improves per-turn token usage tracking, showing input, output, cached, reasoning tokens, and model call counts for each conversation turn.