feat: resolve @mentions in direct reply to Feishu at tags

[lishuceo]

Summary

• 新增 mention-resolver.ts：解析 bot 输出文本中的 @姓名，通过群成员列表匹配 open_id，转换为飞书 post 格式的结构化 <at> 标签
• 在 client.ts 新增 replyPost() / replyPostInThread() 方法
• 修改 sendDirectReply() 短文本分支：检测到 @mention 时用 post 格式发送，无匹配时 fallback 纯文本
• 按姓名长度降序贪婪匹配，避免 @张三丰 被 @张三 先截断
• 任何环节失败（API 调不通、名字不匹配）均 fallback 到原有纯文本逻辑，不会 regression

Test plan

• 11 个单元测试覆盖 convertTextWithMentions 纯函数
• 完整回归测试 1074/1075 通过（唯一失败为已有 flaky test memory/quality）

🤖 Generated with Claude Code

[claude]

Claude finished @lishuceo's task in 1m 21s —— View job

---

Code Review: PR #155 — Resolve @mentions in direct reply

• Check for previous review comments
• Read and review all changed files
• Post review findings

---

✅ Approved — Clean, well-structured PR

Overall assessment: This is a solid implementation. The code follows existing patterns in the codebase, has proper error handling with graceful fallback to plain text at every failure point, and ships with good test coverage (11 unit tests on the pure function).

What works well:

• Greedy longest-match-first strategy via name length sorting prevents @张三 from eating @张三丰
• convertTextWithMentions is a pure function, cleanly separated from the async resolveMentions wrapper — easy to test
• replyPost() / replyPostInThread() follow the exact same pattern as existing client methods
• Regex special characters in names are properly escaped (line 61 of mention-resolver.ts)
• Every failure path (API errors, no matches, empty member list) falls back to the original plain text behavior — zero regression risk

Minor observations (not blocking):

• getChatMembers() paginates through all group members on every @-containing reply. For large groups this adds latency. A short-lived cache (e.g., 60s TTL per chatId) could help if this becomes a bottleneck in practice — but fine to defer.
• Test case on line 88-96 (mention-resolver.test.ts) documents that @张三丰 with only 张三 in the map will partially match, leaving 丰 as trailing text. The comment acknowledges this as acceptable, which is fair — in practice both names would be in the member list.

No issues with confidence ≥ 75 found. LGTM.