日服关键词全渠道检索

三渠道并行搜索指定关键词：

• 玩友圈内容原文 — 数数 SelfContentAudit 事件，包含主帖与评论/回复中的关键词提及
• 私聊消息 — MCP query-user-chat-list + query-user-chat-content
• 家族群 + 语音房公屏 — MCP query-room-chat-log；默认查活动/话题时间段内日服全站活跃家族群 Top10 + 语音房 Top10，不从玩友圈命中用户反推其所在房/群
• 语音房语音转写 — 数数 SelfContentAudit 中 scene = VOICE_ROOM_VOICE 的 ASR 文本；与语音房文字公屏分开统计，作为辅助证据单列

强制口径规则

Failure guards (minimal)

• On Windows/PowerShell, never pipe raw Japanese/Korean/Chinese keywords through stdin or inline scripts without a Unicode round-trip check. Build non-ASCII keywords inside Python with chr(...)/Unicode escapes, or pass them as normal argv only after confirming the echoed keyword is not replaced by ? or mojibake. If a Japanese keyword unexpectedly returns 0 while adjacent broad terms are high, first suspect command/script encoding, not user silence.
• If SelfContentAudit fails, times out, or returns suspicious 0/?-based results, keep the original search intent and retry with optimized count SQL, date/hour sharding, or keyword-safe scripts. Do not downgrade to MCP-only evidence or write "no discussion" until the main SelfContentAudit route is either recovered or explicitly marked as incomplete.
• When the user says an internal Chinese concept maps to a Japanese user-facing term, search the user-facing term only unless they explicitly ask for the Chinese string. Example: internal "段位" for karaoke rank should search ランク, not 段位.
• 时间路由：查询完全位于 2026-01-04 20:05:28 JST 以后时，优先用数数 SelfContentAudit 做全量关键词召回、计数和趋势，MCP 只补代表性原文与上下文。查询涉及更早历史时，必须用旧事件定位对象后再走 MCP 回查；跨越该时间点时拆成两段查询、分别报告，不得直接视为同一召回口径。
• 历史回查路线：私聊用 SendMsgPrivate 定位后调用 query-user-chat-list + query-user-chat-content；群聊和语音房文字公屏用 SendMsgPublic 获取真实 GID/RID 后调用 query-room-chat-log；玩友圈用 PublishState 获取真实 UID 后调用 query-discover-content。空 ID、参数或权限报错不得写成“历史正文不存在”。
• 渠道隔离：查询 SelfContentAudit 必须过滤 scene：玩友圈 DISCOVER、广场 SQUARE、私聊 PRIVATE_CHAT、群聊 GROUP_CHAT、语音房文字公屏 VOICE_ROOM、ASR VOICE_ROOM_VOICE。各渠道分别统计，不得混合后统称“玩友圈讨论”。