[loopx-agent]

Manual explore finding --measurements-json results already contribute to best/latest feedback, but the next-step decision cards previously required the writeback-result origin tag. An agent could see the score and still be unable to bind that exact result to its existing work.

This change includes measured findings in the existing paged result and comparison path. It reuses current node links, source digests, page replay and followthrough validation. Untagged notes without measurements keep their existing summary/audit path; planner scores, execution admission and metric computation are unchanged.

Validation:

• 23 TypeScript tests passed; the untagged cases fail against the unchanged baseline.
• 57 real CLI, File/SQLite and HTTP tests passed. Four final File/SQLite × manual/tagged cases select an older same-scope best of 3 over the latest score of 0, persist its exact source at normal writeback, and reject a wrong digest without appending state.
• TypeScript control-plane typecheck and bilingual RFC check passed.
• Read-only before/after against an existing authorized source: the same latest native measurements become selectable, with identical quantitative feedback. No model, solver or experiment was started.

Required premerge validation exposed legacy smoke fixtures. Both Explore gate smokes now reuse the existing saved cold-import helper. The gate-scope smoke asserts the current fail-closed canonical-authority requirement instead of expecting a retired display-file fallback. Its existing canonical metadata matrix retains the semantic and malformed-exclusion cases. Baseline failures and successful focused repairs were checked independently. Final full premerge passed (5 direct checks, 10 catalog canaries, 8 risk smokes and the changed-path public boundary check; no manual holds). The existing File/SQLite canonical metadata matrix passed 104 cases. CQR: cqr_24e7db027a0e3215b0f5 for this exact diff.

This is a decision-source projection repair. It does not change the frontend layout or claim that recording a choice proves later execution, retained gain or causal improvement. Original experiment adoption and benefit remain separate evidence requirements.

Exact-head author review: #7108 (review) (COMMENTED approval conclusion).

[loopx-agent]

Reviewer: model_agent | gpt-6.1-sol | OpenAI | runtime_reported | reasoning_effort=max

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

动机

使用 Explore 登记评测结果并为已有任务选择下一步的 Agent 或操作者。

同一问题的新候选得 0 分，较早候选得 3 分；此前手动成绩不能绑定选择，现在可翻页找到旧最佳并将其准确来源记入原任务的计划。

同一条手动量化 finding 可以从结果页进入既有选择卡片，正常写回保留准确来源，错误摘要和缺少任务保留声明仍被拒绝。

本次只修决策证据投影及被预合并触发的旧 smoke，不改变自动排序、模型调用、实验执行或研究页面布局。

真实研究源已核对分数及选择卡片；后续原实验是否实际采用某个计划、保留收益与因果效果仍需独立的后续工作证据。

改动思路

成绩与结论的事实来源早已存在：公开登记命令保存同一 finding，比较器计算同口径最佳和变化，原决策校验读取精确来源。断点只是结果列表按写回来源标签过滤，排除了受支持的手动生产者。本次在这个原有列表加入带合法测量的记录，继续复用原分页、当前任务关联和反证优先规则。既有历史和后续执行的边界没有被移到显示层。

复用一个量化事实来源和已有选择校验，只扩展读取投影；测试准备走共享冷导入，未加入执行策略或第二权威。

当前 PR 交付手动成绩到原任务精确计划写回的可用链路；实际后续执行和研究收益仍沿原 owner 验收。

具体改动

关键代码讲解

1. projectExploreTurnContext 在已有 typed measurements 事实或既有来源标签满足时保留结果；不从正文、分数正负或 confidence 猜测来源和任务状态。未带测量的手动笔记仍沿原摘要与审计读取。
2. comparisonCandidates 继续依据当前可执行 Todo 的显式节点关联，绑定完整来源摘要。它与结果页消费同一集合；扩大结果范围后，旧最佳可经原分页返回并在写回重放时核对。错误摘要、缺少原任务保留声明和外来来源仍拒绝。
3. 现有 CLI 集成用例 扩展成手动/带标签两类生产方式，在 File 和 SQLite 中从3、2、1、0分的相同条件成绩里选择较早最佳，验证正常计划写回和错误摘要无追加。类型层复用原3/4/20分页用例覆盖同样的来源差别。
4. 既有配置 smoke、规划门 smoke 复用共享保存式冷导入，不再维护一套手工初始化；gate-scope smoke 验证当前缺权威拒绝。旧元数据的具体负例已有 canonical_todo_semantics_without_display 的 File/SQLite 覆盖，包含错误绑定、未知执行者、移除的策略和畸形排除项，未因旧显示回退被取消而删除这些语义检查。

规范依据是变更前的 Explore 量化结论与后续选择契约。manual-sourced-results：已实现同一事实来源的手动记录；scoped-comparison：原可比键及收益边界保持；exact-next-step：原精确来源选择与写回已验证；default-off-authority：完整关闭态等价及只读证据态已验证。新增中英文段落解释这些行为，没有引入另一项执行义务。

对主干的风险

主要风险是新增结果挤占首屏或被误认成执行许可。当前分页仍有界，相关反证保留原优先级；最新、最佳、未知和候选引用保持原比较器含义，自动分支分数未改。四组 base/head 成对检查证明关闭态与未测手动笔记的完整输出等价；证据模式显示成绩却不产生规划卡片，错误来源摘要在写回前拒绝。未新增兼容格式、状态机或数据库迁移。

23项类型测试、57项真实 CLI/File/SQLite/HTTP 测试与4项最终按成绩选择的集成用例通过，控制面 TypeScript 类型检查通过。初次预合并失败保留：两条 Explore fixture 和一条旧回退 oracle 在未改动基线同样失败，现已沿共享冷导入和当前 canonical 拒绝规则修复，并补跑现有 canonical 负例矩阵。最终完整预合并门通过；按本 Goal wait_for_ci=false，未查询或等待 CI。公开变更路径扫描通过。真实研究数据只读核对时，分数与反馈逐值相同，没有模型或实验执行。

我的整体评价

long_horizon=preserved，user_experience=improved：原持续工作和来源约束保留，手动量化结果可以被既有选择流程使用。本次是 justified_increment，交付了公开生产入口至精确来源计划写回的一段完整修复；选择已记录仍不能当作后续执行或收益已经发生。

复用一个量化事实来源和已有选择校验，只扩展读取投影；测试准备走共享冷导入，未加入执行策略或第二权威。

当前 PR 交付手动成绩到原任务精确计划写回的可用链路；实际后续执行和研究收益仍沿原 owner 验收。

本次未改动任何渲染组件、样式或页面布局；检查的路径是 CLI 结果投影到既有决策校验，视觉评审不适用。

真实实验后续执行、实际保留收益与因果效果未由本 PR 证明；现有真实来源只读验证没有启动模型或实验。

最终差异54行增加、58行删除，生产改动仅原投影中的7行增加、3行删除。局部列表改名、共享 fixture 复用及移除失效回退预期是本范围内的简化，现有持久事实/协议继续使用。未发现此精确 head 的阻塞问题，同意在明确授权、CQR、required validation 与当前 merge-readiness 独立满足后合并。

English verdict: APPROVE e2fbd30. Manual quantitative findings reach the existing exact-source planning consumer without a false writeback-origin tag. Source, paging, disabled-path and canonical authority checks are preserved; local required validation passed. This author review does not certify later experimental execution or benefit.

[loopx-agent]

Approval conclusion (author-owned PR; formal self-approval is unavailable)

Reviewed exact head: e2fbd30cd0c3702f05861d3f7f571bc3a4836aca.

The change correctly admits measured findings into the existing exact-source decision path without requiring the writeback-result tag. The predicate excludes absent, unmeasured notes, page replay retains complete measured sources, and selection still passes source digest and followthrough validation.

Independent exact-head validation:

• 15 Explore TypeScript tests passed, including all tagged/untagged page sizes.
• 6 real CLI tests passed for manual and writeback-tagged measurements across File and SQLite.
• Control-plane TypeScript typecheck and bilingual RFC check passed.
• Standard premerge passed all 19 selected checks with zero failures and zero manual holds.
• Public-boundary scan passed.

No blocking issue found. CI is still running; this review grants no merge authority and does not claim experiment adoption or long-term benefit.

English verdict: APPROVE — e2fbd30cd0c3702f05861d3f7f571bc3a4836aca preserves the exact-result binding boundaries while making manually measured findings selectable through the existing paged comparison owner.