feat(tools): semantic tool-name suggestion for registry misses

[raymondginger2018-sudo]

Summary

When the LLM hallucinates or misremembers a tool name, the registry returns "not found". This PR adds a cheap, deterministic first mitigation: given the missed name and the available tool names, suggest the closest candidates by token-overlap + sequence similarity.

Design

• Pure mechanism: no LLM calls, no embeddings, no network — uses difflib.SequenceMatcher + token-overlap scoring.
• Safety guard: semantic discovery is only a hint fed back to the model as an error message. Execution still requires the exact registered name plus the permission engine — it never widens the callable surface.
• Zero new dependencies (stdlib only).

Files

• core/agent_runtime/tools/semantic_hint.py (new, 82 lines)

Testing

• Module is side-effect free and importable standalone.
• Callable surface unchanged; no changes to existing registry code.

Inspired by GenAI lesson 17 (P2-A7): Taskweaver-style semantic plugin discovery, adapted as a non-executable hint.

[raymondginger2018-sudo]

设计说明

问题：当 agent 调用工具名拼写有微小偏差（如 execute_bash 写成 exec_bash、read_file 写成 readfile），注册表查找直接返回空，agent 进入"工具不存在"的死胡同。

解法：用 difflib.SequenceMatcher 计算名字相似度 + token 交集双重判定，返回 Top-3 候选提示。pure mechanism，无 I/O，无 LLM 调用。

关键设计决策：

• 阈值 0.6 定得宽松（宁可误报也不漏报），因为结果是 suggestion 而非自动 fallback
• _tokens() 按 snake_case 和 camelCase 分词，但对短名退化到字符级
• 所有函数都没有副作用，纯计算，适合缓存在 runtime 启动时预热

测试建议：提供几个已知 miss case 作为回归测试即可（exec_bash → execute_bash、file_read → read_file）