fix(workspace): show native measurements alongside research judgments

[loopx-agent]

Research judgments were visible while the native measurements they should be checked against were hidden inside the source accordion. This moves the existing source-bound measurements directly below the six analyzer metrics: initial, best within the captured window, provider-retained and final valid submission stay distinct. Values retain their exact precision, direction, window and coverage; an unknown initial, absent capture or changed source remains explicit.

The six analyzer metrics remain primary, costs remain folded, and no gain, adoption decision or combined research score is inferred. Reading values now takes zero disclosure clicks; the original cited source is one optional click, with keyboard focus restored on return. Candidate identities and submit/receive times remain available in measurement details. Browsing and source selection do not request analysis.

Placement: this only formats the existing benchmark-toolkit projection and citation callback in the Goal workspace. No producer, API, schema, configuration, CLI/Lark entry or execution authority changes. The bounded refactor uses one local native readout and removes the previous inline duplicate. Scoped styles preserve 44px evidence targets despite the Goal's shared button rules.

Visual evidence

Packaged browser UI on the repository's populated synthetic showcase; every metric/score in these public images is invented. The before frontend is f9bdae67596c185cebc81dc555dd6a3e0242e9c5, whose affected frontend sources are unchanged at base 8795fc042550c875d78b8bfc1a4874523fbb6693. The after source is 3be13959c0bdd3d0b85286622ecc23ded020d78b. These are browser captures, not an installed App, a live model or scientific qualification.

The whole Goal first viewport remains the same composition. The research section sits below it; the comparisons show that actual reading position rather than implying that every region fits above the fold.

Actual first viewport: desktop and narrow

Research reading view	Before	After
Desktop, research section
390px, native values

The narrow before image has the source disclosure opened; the after image needs no such disclosure. Both are scrolled views. Author-local read-only synthetic preview is ephemeral and is not a hosted deployment.

Validation

• Final-head TypeScript/package build, source-manifest verification and CSS token check pass. The existing large-chunk warning remains.
• Nine actual HTTP/store/process tests pass, including source binding, wrong Goal/origin rejection, native facts before analysis and immutable citation return. Synthetic provider output is not model calibration.
• Ego space 247: desktop and 390px, Chinese/English, reduced motion, exact 0.000001, distinct best/retained/final, partial/missing/historical captures, source switching and refresh, keyboard source opening and return focus pass. Scoped evidence buttons measure 44px; no horizontal overflow. Request readback contains no analysis POST.
• Existing completed real Graph judgments were read without rerunning analysis; a separately authorized timed Portfolio result was read through the same actual owner and citation route. Its missing initial and partial coverage stay unknown. Private evidence and real run data are excluded from this PR.
• Updated the durable research-history scenario for direct values, stale/missing scopes, retained-versus-final and partial prefixes. Its full automated browser catalog was not run locally; the affected journey was exercised in Ego. CI remains a separate gate.
• Full semantic smoke, DCO, bilingual-diff and public-boundary checks pass. Premerge passes all four direct and thirteen selected checks against target snapshot c7061d7acec7e87b049a567788b8070bde4895fd, with exact-scope receipt cqr_64df6db5f7c152cf005e. A later target change requires requalification before merge. The advisory excludes TSX; the native point vocabulary was manually verified as existing local display keys.

此改动让六项分析判读与同范围原生测量分别可读，不把“检验过收益”或“证据采用”说成收益改善。截图为合成数据；真实案例另做本机只读验收。当前仅提案，尚未合并或正式安装。

[loopx-agent]

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

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

APPROVE；本轮未发现阻塞项。评审绑定完整 head 3be13959c0bdd3d0b85286622ecc23ded020d78b，基线 8795fc042550c875d78b8bfc1a4874523fbb6693；批准限定下述已验证变化，剩余资格不记为通过。

动机

研究 Goal 的读者需要在六项分析判读旁边直接看到同一范围的原生测量，并能打开数值所引用的原文。

旧界面把已有原生值藏在来源折叠区中，读者先展开才能核对结果。候选将首次、窗口最佳、提供方保留和最后有效值直接排在六项判读下方；阅读数值不再需要展开来源。

同一真实存储/HTTP 读回配合基线和候选打包前端，验证了0.000001原精度、7/0.000001/7/5四类不同值、历史标记、缺失不借用、部分窗口首次未知，以及键盘打开原文后返回数值焦点；没有分析 POST。

此显示改动不产生新评分、收益推断、证据采用或分析执行，也不改变采集器；科学校准、全研究成本、真实用户理解效果和正式安装不由本 PR 完成。

评审规范：docs/development/design.md，不可变版本 8795fc042550c875d78b8bfc1a4874523fbb6693。逐项对照：

• Earn The User's Attention：Start from the owner task, avoid redundant preparation and distinguish unknown observations from reassuring success. 由本轮对应真实入口与负例验证；更广资格仍按下文边界保留。
• Responsive Behavior：Reflow narrow screens while preserving readable evidence and keyboard access. 由本轮对应真实入口与负例验证；更广资格仍按下文边界保留。
• Source Of Truth：Preserve behavior, accessibility and public/private boundaries through existing owners. 由本轮对应真实入口与负例验证；更广资格仍按下文边界保留。

改动思路

一个局部 NativeResults 只格式化既有 typed 原生投影，沿用原引用回调，并删除旧的重复展示；判断与采集继续归现有 owner。

当前 PR 边界是 Goal 工作区的只读数值呈现、局部样式、对应浏览器场景及公开合成图片。

保持现状会继续隐藏决定性事实；增加摘要分数或重算收益会制造第二个判断 owner；仅复制原折叠内容又会产生两份维护点。移动并局部提取现有显示，是更小且完整的阅读改进。

同一捕获的事实	读回
首次 / 窗口最佳 / 原生保留 / 最后有效	7 / 0.000001 / 7 / 5，未合并为一个“收益”
来源变化 / 无捕获 / 部分前缀	历史且当前未核验 / 明确缺失 / 首次未知而非零
查看原文	一次可选操作；Escape 返回原数值焦点；无需发起分析

原生测量、未校准的模型判读与实际收益是三种不同证据。费用仍保持自己的窗口、单位与未知边界。

具体改动

• NativeResults：Format existing native outcomes without deriving utility; preserve current/historical scope, exact numeric string and unknown entries.
• ResearchProgress：Keep six analyzer metrics primary; place native values before folded source/cost details using the existing citation callback.
• .research-native-values / scoped native buttons：Four columns become two on narrow screens; evidence hit targets stay44px despite shared Goal button rules.

NativeResults 读取现有 native_outcomes.projection，用已有四个键、原值字符串和原引用显示观测；历史来源有明确警告，缺失来源独立显示未知。候选没有用分析器的“收益检验”或“证据采用”来补原生得分，也没有把最后提交当作保留结果。

旧来源区里的重复原生显示被移除；候选身份、提交/收讫时间、原始限制仍在可选详情中。局部CSS负责四列/两列重排及44px引用按钮；research-history.mjs 更新直接可读、历史、缺失、部分窗口和保留/最后不同的断言。六张图片为明确标注的合成展示，首屏、桌面与390px前后图在PR正文，未借它们声称已安装或有科学收益。

对主干的风险

最强反例是：Native file fingerprint change returns historical measurements; merely changing a different assessment source does not automatically invalidate the native artifact. Missing capture has no previous-scope values. Partial prefix has unknown initial but retains separately valid best/retained/final. Reload preserves historical scope; citation before an assessment works without executing analysis.

3 actual HTTP/process tests plus12 store tests passed independently (15 total); TypeScript, Chat package build, source-manifest verify and CSS tokens passed. Standard premerge4 direct+13 selected passed. Ego281 covered paired source visibility, scope changes, partial/missing/history, exact precision, citation/return focus,390px English/Chinese and reduced motion. All four evidence targets measured44px; no horizontal overflow.

Build retained its pre-existing large-chunk warning. Early browser snapshot used an unsupported CSS-root option; data/actions had completed and a valid snapshot followed. An initial screenshot did not include the lower mobile native section; an explicit native-section capture superseded that framing. The source-specific native invalidation fixture was corrected rather than treating an unchanged native file as stale.

未查询、轮询或等待 CI；本结论基于当前本地源码、真实受影响入口与明确隔离的合成/原生对照。源码提案、打包界面、已安装产品与真实模型执行分别判断。

我的整体评价

当前边界有直接用户价值，替代方案和相邻 owner 的检查没有发现必须扩大范围或新增第二套规则的理由。未来维护所需的小重构已体现在复用既有 owner 与局部提取中；未引入纯为未来准备的框架。一个局部 NativeResults 只格式化既有 typed 原生投影，沿用原引用回调，并删除旧的重复展示；判断与采集继续归现有 owner。 当前 PR 边界是 Goal 工作区的只读数值呈现、局部样式、对应浏览器场景及公开合成图片。

No blocking finding. Synthetic provider results do not establish calibrated model judgments, native study gains or whole-study cost. The complete research-history browser catalog was not rerun independently; affected packaged reading and real HTTP/store/process paths were exercised. No installed App, paid model, new benchmark or causal benefit is claimed.

这是共享作者账号上的 COMMENTED 批准结论，不是 GitHub 正式自批准。旧阻塞、线程状态和合并 readiness 仍须分别回读；本评审不授予合并、安装或执行权限。

English verdict: APPROVE — exact head 3be1395. Native measurements are directly readable through the existing typed projection. 15 backend tests and paired packaged UI checks passed for precision, missing/history scope, citations and narrow screens. Synthetic evidence establishes no study gain or total cost; no merge or install authority is granted.