docs(steward): distinguish source deduplication from unfinished work

[huangruiteng]

Goal And Delivered Outcome

GQ06 asks, “Read this. Save what's useful and follow through on improvements.” Its existing fixture distinguished an indexed copy from a new revision, but could reward skipping unfinished work or creating duplicate work after a spelling-only change.

This defines the material-to-work pilot in the existing golden-query pack, linked from the bilingual M2/M3 checkpoint. Source identity, contextual relevance and an independently satisfied outcome are judged separately. Base: main; existing roadmap R3 and semantic-handoff M2/M3.

Scope And Continuation

• Seven fixed counterfactuals cover supported no-change, an indexed but unfinished proposal, irrelevant revisions, relevant deltas into existing work, responsibility-based routing, incomplete evidence and instructions quoted in source material.
• Assessment, implementation, validation and return remain distinct from supplied/read receipts. Correction and lost-acknowledgement replay must retain the original request without duplicate effects.
• This is a specification and qualification prerequisite, not an implemented or installed GQ06 improvement. The next existing GQ06 slice executes the App journey and observes actual receiver assessment and original-conversation return on a pinned release candidate. Paid semantic/model evaluation stays release-only; Lark acceptance is separate.
• Reuses existing conversation/artifact and collaboration/work-item owners, with TypeScript decision/effect ownership. No new store, classifier, inbox, capability, mandatory provider or test scaffolding. Ordinary reading does not activate Material Lifecycle.

Validation

• Tested revision: 226284ed0acf8d3ccdb2abf8d8c3420eaa0b7ae5
• Run state: finished
• Input classes: synthetic, public_fixture

Check kind	Result	Evidence / limitation
static	passed	git diff --check origin/main...HEAD; three changed documents scanned for private material; 62 local links and three GQ06 anchors resolve.
manual	passed	Seven counterfactuals compared to independently specified facts; EN/ZH constraints, ownership and unqualified status match.
real_entrypoint	passed	uv run --extra test loopx canary premerge --from-git-diff --no-progress: 18 selected checks and three direct checks passed; no failures, skips, manual holds or tracked side effects. Exact diff change-quality receipt verified valid.
real_backend	not_applicable	No runtime, state, provider or persistence changes. No live bot installation, dispatch or receiver adoption claimed.
manual	not_run	No rendered Markdown or installed App/Lark/model journey evidence; those are not claimed as product acceptance.

Coverage applies to this documentation diff. The native checks exercise their disposable fixtures; they do not qualify the new semantic journey.

Frontend / Visual Evidence

UI impact: none. No shipped UI, first viewport, navigation, prompt or configuration changes. The existing Markdown case pack is extended without claiming rendered or packaged presentation validation.

Type of Change

• Documentation update

LoopX Area

• Public docs or presentation surface

Technical Direction

R3 responsible dispatch and semantic-handoff M2/M3. Future-facing pass consolidates the oracle in one existing pack and links it from both RFC languages; runtime ownership remains unchanged.

Shared-authority RFC fixture impact

N/A: this does not claim TypeScript migration or authority promotion. Existing typed owners are preserved as the planned implementation boundary.

Boundary Checklist

• No private state, credentials, raw traces, internal links or local paths in the diff or public artifacts.
• No duplicate maintainer-owned benchmark work.
• Scoped to the existing GQ06 qualification gap.
• UI impact marked none.
• DCO sign-off included.

[huangruiteng]

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

审阅 head：226284ed0acf8d3ccdb2abf8d8c3420eaa0b7ae5。未发现阻塞此文档增量的问题；这不是 GQ06 产品验收通过或合并授权。

动机

GQ06 的旧例只区分“已收录”和“新版本”，没有固定实际工作是否完成。来源相同却仍有待修问题，或来源只改了拼写，都可能被错误评分。这次将来源去重、当前适用性和结果验证分别比较，形成可用于下一轮验收的有用前置规格。用户仍只说普通文章请求，无须提供 Agent、digest 或工作流名称。产品闭环与持续推进仍待实际执行证明。

改动思路

沿既有 M2/M3 拆清输入、获授权知识写入、负责人评估、执行及原入口回传，复用 conversation/artifact 与 collaboration/work-items。现有 presentation RFC 已引用 Botmux；再次做专用集成或材料存储没有必要。保持共享决策在 TypeScript owner，资料阅读不激活 Material Lifecycle，不靠新的 inbox、分类器或 runtime 状态承载这一试点。

具体改动

关键内容讲解

golden-queries.md 替换原先笼统的 twin/revision 段落，集中描述七种固定事实：已完成的不改、已收录但未完成的推进、无语义增量的新版本、有增量的已有工作、按职责找到负责人、资料不完整及资料内越权指令。评估回执、代码实现、独立验证和结果回传分别观测；同一 Agent 能直接处理时不强制委派。

英文和中文 semantic-handoff RFC 链接同一验收依据，并明确普通阅读、来源证据、执行权限的边界。纠偏、丢失 ACK、刷新与重放要保留约束和原请求关系；真正改变的问题重新评估。没有新增自动加载指令、默认行为、schema 或强制 provider。

对主干的风险

最大风险是把写好的评分说明当成已经可用的产品，或把“已收录”再次误当成实际完成。三份文档都保持未验收状态，具体要求负责人评估和实际行为证据；Botmux 研究也限制在固定源码/版本的观察，未声称通用保证。

本地 exact-head 原生 premerge 的 18 个选中检查及 3 个直接检查通过；无失败、跳过、manual hold 或 tracked side effect。三文件公开边界干净，62 个本地链接和 3 个段落锚点有效，EN/ZH 语义人工对照通过，change-quality 收据匹配当前 diff。未拉取或等待 CI。没有渲染 Markdown、付费模型或安装态 App/Lark 实验，不能据此断言材料被接收方采用。

我的整体评价

这是合理的验收准备增量：它排除了两个具体的错误评分方向，并把详细 oracle 留在单一既有文档；代码兼容路径、actor 生命周期、权限与调度均未改变。对持续推进和用户旅程的运行行为保持原状，下一步仍由既有 GQ06 工作在 release candidate 上验证 App 的真实接收、纠偏、恢复和一次原会话回传。最值得的简化已经应用：替换旧段落、RFC 只引用同一依据，不增加临时 smoke 或第二套材料能力。保留给维护者合并。

English verdict: APPROVE - HEAD 226284e. Useful GQ06 qualification prerequisite distinguishes indexed knowledge from satisfied work. Static/bilingual review and native premerge passed; no shipped semantic or installed-channel acceptance is claimed.