fix: stop @_all from triggering bot mention; coerce MCP enum types

[Keithin]

What

Two fixes in this PR:

1. @_all false-positive fix — Removed the @_all short-circuit in the Feishu gateway adapter that was causing bot mention detection to fire on @_all messages. Now only explicit personal @ mentions trigger bot reply.

2. MCP enum type coercion — Added _coerce_enum_types() in mcp_tool.py to handle MCP tool schemas where enum values are strings (e.g. "10") but the declared type is integer. Oceanengine MCP tools send schemas like {enum: ["10","20"], type: integer}; this fix recursively coerces enum values to match the declared type.

Testing

• @ALL -> silence ✅
• @Hermes -> reply ✅
• @ALL @Hermes -> reply ✅
• Single chat -> always reply ✅

[alt-glitch]

This was generated by AI during triage.

Related: the Feishu @_all hunk overlaps open #59347, while this PR also carries a distinct MCP enum-schema coercion fix. The bundled scopes should be reviewed independently.

[teknium1]

Thanks for addressing two current-main defects. The Feishu premise is confirmed at plugins/platforms/feishu/adapter.py:4399-4413, where raw @_all currently returns a self-mention; MCP normalization likewise has no enum coercion at tools/mcp_tool.py:5330-5340.

Problems

• The new user_status_change registration will make the exact expected call list fail in tests/gateway/test_feishu.py:434-450; add customized:user_status_change before build.
• tests/gateway/test_feishu.py:1664-1677 currently asserts the opposite @_all behavior. Update it to cover @_all alone being rejected under mention gating, while explicit bot mentions still pass.
• Add MCP regression coverage: current schema tests cover definitions and nullable unions (tests/tools/test_mcp_tool.py:290-353), not recursive integer/number enum coercion.

The member note on #72607 correctly identifies the Feishu overlap with #59347 and the MCP repair as distinct scopes; preserve test coverage for both when salvaging. This is an automated hermes-sweeper review.

[teknium1]

Please update TestAdapterBehavior.test_build_event_handler_registers_reaction_and_card_processors: its exact expected call list at tests/gateway/test_feishu.py:434-450 currently omits this registration, so this hunk will fail that test.

[Keithin]

补充复现证据（2026-08-14，v0.20.1 仍受影响）：

影响持续：此 bug 已导致我们本地第 5 次回归（hermes update 每次 reset 后 _mentions_self 的 @_all 短路都会恢复）。每次升级后群聊 @所有人 都会被机器人回复，需要手动重打补丁。

复现步骤（v0.20.1 / origin/main 当前代码）：

1. 机器人加入飞书群
2. 群成员发一条只 @所有人（不 @机器人）的消息
3. 观察：_admit → _mentions_self 命中 if "@_all" in raw_content: return True → 机器人错误回复

根因定位：plugins/platforms/feishu/adapter.py _mentions_self() 的 @_all 短路把 @everyone 当成 @bot。修复 = 删除短路，只认 mentions 数组里 bot open_id（本 PR 已实现，本地验证通过：@所有人 静默、单独 @bot 回复、@bot+@ALL 回复）。

附加发现（供参考）：上游 _process_inbound_message 对"纯 @bot 无正文"消息会丢弃（Ignoring empty text message），但这对分段指令/上下文提醒场景是有意义的唤醒信号。我们已本地改为：mentions 含 bot 时放行（not self._message_mentions_bot(mentions) 守卫）。如需可另开 PR。

PR 修复经过本地 2 周生产验证，恳请维护者 review/merge。