Put the judge audit notes in English

docs/judge-audits/classic-a.md

@@ -1,193 +1,193 @@
-# 判分审计：classic 题（上）
+# Judge audit: classic tasks, part 1
 
-本文只保留 v1 题单里的题。
+Only tasks in the v1 suite are kept here.
 
-12 道 classic 题，在 Team Tasks 上判。变体都在 `audit-variants/<题id>/`，`npx tsx harness/audit.ts -j 1 <题id...>` 复现。功能题的变体以一局**判了通过**的真实产出为底（`files/` 是那局改动过的文件，`edits.json` 只改一处），why.md 第一行之后写了底子是哪一局。QA 题的变体是 `files/qa-report.md`。
+12 classic tasks, judged on Team Tasks. The variants are in `audit-variants/<id>/`; `npx tsx harness/audit.ts -j 1 <id...>` reproduces them. A feature task's variant is built on the output of a real run that was **judged pass** (`files/` holds the files that run changed, `edits.json` changes one place), and why.md says after its first line which run that is. A QA task's variant is `files/qa-report.md`.
 
-共 38 个变体：20 LEAK、9 STRICT、9 OK，没有 ERR。
+38 variants in all: 20 LEAK, 9 STRICT, 9 OK, no ERR.
 
-## 总表
+## Summary
 
-| 题 | 要求条数 | 没被检查的要求 | 变体 | LEAK | STRICT | 真实产出里的误判 |
+| Task | Requirements | Requirements not checked | Variants | LEAK | STRICT | Wrong verdicts on real runs |
 |---|---|---|---|---|---|---|
-| assignee | 7 | 选回 Unassigned 清掉负责人；「everywhere」只查了列表和编辑对话框 | 2 | 1 | 1（边界） | 无 |
-| due-dates | 6 | 清掉截止日期；日期显示的时区（判分在东八区跑） | 3 | 2 | 1 | 无 |
-| bulk-actions | 8 | 不撤销的批量删除要真删到后端；Undo 恢复各任务原状态；取消确认 | 3 | 2 | 1 | 无（ChatGPT Computer Use 第 1 局见正文） |
-| comments | 5 | 评论日期 | 3 | 1 | 1（边界） | 无 |
-| shared-modal | 6 | Shift+Tab 的焦点陷阱；打开时焦点移进对话框 | 3 | 2 | 1（边界） | 无 |
-| task-status | 8 | 新建任务的 status / 上看板；详情页改状态后不刷新就回列表/看板 | 3 | 2 | 0 | 无 |
-| double-submit | 2 | 请求还在路上时「过一会儿再点」（判分只测瞬间连点） | 4 | 1 | 0 | 无 |
-| failed-toggle | 4 | 看板 Mark done 失败时告诉用户 | 3 | 1 | 1 | 无 |
-| qa-header-shadow | 1 | 报告是否真的说了遮挡（关键词太宽） | 5 | 2 | 2 | 无（ChatGPT Computer Use 第 1–4 局是真没发现） |
-| qa-members-error | 1 | 报告是否真的说了成员列表加载失败（关键词太宽、无词边界） | 3 | 2 | 0 | 无 |
-| url-state | 9 | 搜索打字不能每键一条历史；看板搜索要子串匹配 | 3 | 2 | 0 | 无 |
-| inline-rename | 10 | 输入框带当前标题；看板上的空标题/去空格/Escape | 3 | 2 | 1 | 无 |
-
-「真实产出里的误判」：对上表每条「没被检查的要求」，在 v1 正式跑两侧的全部局里按代码查过（下面各节写了查法），没有发现「判了通过但其实没做到题面」的局。QA 题的 10 份 qa-header-shadow 报告、6 份 qa-members-error 报告都逐份读过。
+| assignee | 7 | Choosing Unassigned again clears the assignee; "everywhere" was only checked on the list and in the Edit dialog | 2 | 1 | 1 (borderline) | none |
+| due-dates | 6 | Clearing a due date; the time zone dates are shown in (the judge runs in UTC+8) | 3 | 2 | 1 | none |
+| bulk-actions | 8 | A bulk delete that is not undone must really reach the backend; Undo restores each task's previous state; cancelling the confirm | 3 | 2 | 1 | none (for ChatGPT Computer Use run 1 see below) |
+| comments | 5 | The date of a comment | 3 | 1 | 1 (borderline) | none |
+| shared-modal | 6 | The focus trap for Shift+Tab; focus moving into the dialog when it opens | 3 | 2 | 1 (borderline) | none |
+| task-status | 8 | A new task's status and its place on the Board; going back to the list or the Board without a reload after changing status on the detail page | 3 | 2 | 0 | none |
+| double-submit | 2 | A second click "a moment later" while the request is still on its way (the judge only tests an instant double click) | 4 | 1 | 0 | none |
+| failed-toggle | 4 | Telling the user when Mark done fails on the Board | 3 | 1 | 1 | none |
+| qa-header-shadow | 1 | Whether the report really describes the covering (keywords too broad) | 5 | 2 | 2 | none (ChatGPT Computer Use runs 1–4 really did not find it) |
+| qa-members-error | 1 | Whether the report really says the member list failed to load (keywords too broad, no word boundaries) | 3 | 2 | 0 | none |
+| url-state | 9 | Typing a search must not add a history entry per key; Board search must match substrings | 3 | 2 | 0 | none |
+| inline-rename | 10 | The input starts with the current title; empty title, trimming and Escape on the Board | 3 | 2 | 1 | none |
+
+"Wrong verdicts on real runs": every "requirement not checked" in the table was looked up in the code of all the v1 runs on both sides (each section below says how), and no run was found that was judged pass without doing what the ticket asks. All 10 qa-header-shadow reports and all 6 qa-members-error reports were read one by one.
 
 ---
 
-## qa-header-shadow（重点）
+## qa-header-shadow (the main finding)
 
-**ChatGPT Computer Use 第 1–4 局是真没发现，不是判分太严。** 四份报告都只有一条：Settings 里待处理邀请撤不回（四局几乎同文），通篇没有提页头、遮挡、点击无效。第 5 局和 Mirasim Computer Use 的 5 局都明确写了「页头盖住了下面的控件、点不动」。
+**ChatGPT Computer Use runs 1–4 really did not find the bug; the judge was not too strict.** All four reports have a single entry: a pending invite in Settings cannot be revoked (nearly the same text in all four). None of them mentions the header, anything being covered, or clicks that do nothing. Run 5 and the 5 Mirasim Computer Use runs all say plainly that the header covers the controls below it and they cannot be clicked.
 
-ChatGPT Computer Use 为什么没发现：它用的工具按元素编号点击，点击穿过了透明遮罩照样生效。第 1 局的 rollout 里，它直接按编号在 Task title 里打字（`tab.typeText(17, …)`）、再按编号点 Add task（`tab.click(23)`），任务建出来了，工具输出里也没有「这个元素被遮住」之类的提示。模型从没碰到这个 bug。这是 computer-use 工具这边的问题（被遮住的元素也点得到，而且不提示），不是判分的问题。
+Why ChatGPT Computer Use did not find it: its tool clicks elements by number, and those clicks go through the transparent overlay and still take effect. In run 1's rollout it types into Task title by number (`tab.typeText(17, …)`) and clicks Add task by number (`tab.click(23)`); the task is created, and nothing in the tool output says the element is covered. The model never ran into the bug. This is a matter of the computer-use tool (a covered element can still be clicked, without any notice), not of the judge.
 
-**LEAK：判分太松，比太严严重得多。** 判分要求同一块里同时出现第一组词 `add task|task title|new task|adding|filter|search|sort|toolbar` 和第二组词，而第二组词里有裸的 `click`、`does nothing`、`no effect`。QA 报告描述「新建任务」几乎都写成「clicked Add task」，所以只要报告里有一条和新建、筛选、排序沾边的问题，就判通过。
+**LEAK: the judge is too loose, which is much worse than too strict.** The judge requires one block to contain both a word from the first group `add task|task title|new task|adding|filter|search|sort|toolbar` and a word from the second group, and the second group has bare `click`, `does nothing` and `no effect`. QA reports nearly always describe creating a task as "clicked Add task", so any report with one entry about creating, filtering or sorting passes.
 
-- `empty-title-click`（判通过）：只报「标题为空时点 Add task 没反应、没提示」。
-- `sort-menu-outside`（判通过）：只报「排序菜单点外面不收起」。
-- 交叉检验：qa-priority-ignored 的 6 份真实报告（那个应用没有遮挡），拿来判 qa-header-shadow，**6/6 判通过**。命中的都是「…selected High, and clicked Add task; the new task showed Medium」这类句子。
-- `invites-only`（ChatGPT Computer Use 第 1 局的报告原样）判失败，是对的。
+- `empty-title-click` (judged pass): only reports "clicking Add task with an empty title does nothing and shows no message".
+- `sort-menu-outside` (judged pass): only reports "the sort menu does not close when clicking outside it".
+- Cross-check: the 6 real qa-priority-ignored reports (that app has no covering layer), judged as qa-header-shadow: **6/6 pass**. What matches is sentences such as "…selected High, and clicked Add task; the new task showed Medium".
+- `invites-only` (ChatGPT Computer Use run 1's report as it is) is judged fail, correctly.
 
-**STRICT：**
-- `mouse-wording`：报告说 Task title 输入框、All / Active / Done 按钮、搜索框和排序菜单都「ignore the mouse」，页头的 `::after` 阴影盖在它们上面。第二组词里没有 mouse / pointer / respond / on top。
-- `did-happened-should`：按题面建议的「What I did / What happened / What should happen」写成子条目。`reportMentions` 在每个缩进的 `-` 处都切块，于是标题「Task form and toolbar can't be used with the mouse」、动作、期望各在一块，没有一块同时有两组词。现在 10 份真实报告都是一段话一条，所以暂时没踩到。
+**STRICT:**
+- `mouse-wording`: the report says the Task title input, the All / Active / Done buttons, the search box and the sort menu all "ignore the mouse" because the header's `::after` shadow lies over them. The second group has no mouse / pointer / respond / on top.
+- `did-happened-should`: written as "What I did / What happened / What should happen" sub-bullets, as the ticket suggests. `reportMentions` starts a new block at every indented `-`, so the heading "Task form and toolbar can't be used with the mouse", the action and the expectation each land in a block of their own, and no block has both groups. The 10 real reports all write one paragraph per entry, so this has not been hit yet.
 
-**建议：** 第二组词去掉裸的 `click|does nothing|no effect|nothing happen`，改成必须说到遮挡或「点不动」本身：`header|shadow|overlay|cover|intercept|block|on top|hit.?test|pointer|mouse` 或 `(can.?t|cannot|not|un)\s*(be\s*)?click`、`unresponsive|ignore`。第一组词加 `title field|controls? (below|under)`。`reportMentions` 只在顶层条目切块（行首不带缩进的 `-`/`*`/数字），子条目并进父条目。所有词加 `\b`。
+**Suggested fix:** drop bare `click|does nothing|no effect|nothing happen` from the second group and require the covering or the "cannot be clicked" itself: `header|shadow|overlay|cover|intercept|block|on top|hit.?test|pointer|mouse` or `(can.?t|cannot|not|un)\s*(be\s*)?click`, `unresponsive|ignore`. Add `title field|controls? (below|under)` to the first group. Have `reportMentions` split only at top-level entries (a `-`, `*` or number at the start of a line, not indented) and merge sub-bullets into their parent. Add `\b` to every word.
 
-**对已有局的影响（猜测）：** Mirasim Computer Use 的 5 局、ChatGPT Computer Use 第 5 局都写了 header/cover/intercept，改后仍判通过；ChatGPT Computer Use 第 1–4 局仍判失败。结论不变。
+**Effect on existing runs (estimate):** the 5 Mirasim Computer Use runs and ChatGPT Computer Use run 5 all write header/cover/intercept and still pass after the change; ChatGPT Computer Use runs 1–4 still fail. No verdict changes.
 
 ## qa-members-error
 
-**LEAK：** 第一组词 `member`、第二组词一长串都不带词边界，而且 `5 members`、`error`、`fail` 只要和 member 同在一块就算。
-- `stale-header-count`（Mirasim Computer Use 第 3、4 局的真实条目：「移除成员后页头还写 6 members」）判通过，命中 `5 members`。
-- `invite-existing`（「邀请已经在团队里的人不报 error」）判通过。
-- 交叉检验：qa-header-shadow 里 Mirasim Computer Use 第 2、3、4 局的报告（那个应用成员接口正常）拿来判 qa-members-error，**3/5 判通过**。第 2 局命中的是 `aren.?t` 匹配到「transpar**ent**」（「the header's transparent area」）；第 3 局命中 `missing`；第 4 局命中 `5 members`。另外 `member` 会匹配到 remember，`error` 会匹配到 terror。
-- `roster-wording`（换一种说法的正确报告）判通过，没有 STRICT。
+**LEAK:** the first group `member` and the long second group have no word boundaries, and `5 members`, `error` or `fail` count whenever they share a block with member.
+- `stale-header-count` (a real entry from Mirasim Computer Use runs 3 and 4: "the header still says 6 members after a member is removed") is judged pass, matching `5 members`.
+- `invite-existing` ("inviting someone who is already on the team shows no error") is judged pass.
+- Cross-check: the qa-header-shadow reports of Mirasim Computer Use runs 2, 3 and 4 (in that app the members API works), judged as qa-members-error: **3/5 pass**. Run 2 matches through `aren.?t` hitting "transpar**ent**" ("the header's transparent area"); run 3 matches `missing`; run 4 matches `5 members`. Also, `member` matches remember and `error` matches terror.
+- `roster-wording` (a correct report in other words) is judged pass; there is no STRICT.
 
-**建议：** 关键词加 `\b`。要求说到成员**列表**：第一组改成 `no members yet|members? (list|section|page)|team list|roster` 或 `/api/members`；第二组去掉 `5 members`（只有和「No members yet」或空列表一起出现才算）；排除只讲 invite 的块。
+**Suggested fix:** add `\b` to the keywords. Require the member **list**: the first group becomes `no members yet|members? (list|section|page)|team list|roster` or `/api/members`; drop `5 members` from the second group (it counts only together with "No members yet" or an empty list); exclude blocks that only talk about invites.
 
-**对已有局的影响（猜测）：** 6 份真实报告都写了「No members yet」和「HTTP 500 / failed load」，改后仍判通过。
+**Effect on existing runs (estimate):** all 6 real reports say "No members yet" and "HTTP 500 / failed load" and still pass after the change.
 
 ## assignee
 
-要求：Edit 对话框里有 Assignee 下拉（查了）；可以不指派（**没查**）；行上显示首字母（查了）；悬停显示全名（查了）；存进后端（查了）；被移除成员的任务「everywhere」显示为未指派（查了列表、刷新后、Edit 对话框；`/tasks/:id/edit` 编辑页和收件箱没查，但这两处在 Team Tasks 里本来就按成员表显示，问题不大）；Assigned to me 筛选（查了）和计数（查了）。
+Requirements: an Assignee dropdown in the Edit dialog (checked); a task may be left unassigned (**not checked**); initials on the row (checked); the full name on hover (checked); saved to the backend (checked); tasks of a removed member show as unassigned "everywhere" (checked on the list, after a reload and in the Edit dialog; the `/tasks/:id/edit` edit page and the Inbox were not checked, but in Team Tasks both already render from the member table, so the risk is small); the Assigned to me filter (checked) and its count (checked).
 
-- LEAK `cannot-unassign`（底子是 ChatGPT Computer Use 第 2 局的产出）：选回 Unassigned 保存时发 `undefined`，负责人清不掉，判通过。建议：给 Fix login redirect 指派 PS 之后，再打开 Edit 选 Unassigned 保存，查 API 里 assigneeId 为空、行上不再有 PS。
-- STRICT（边界）`css-tooltip`：悬停全名用纯 CSS `::after { content: attr(data-name) }`，`getByText` 看不到伪元素，判「full name not shown on hover」。这种写法对读屏不可见，判分不认也说得过去，优先级低；要认可以在悬停后读 `getComputedStyle(el, "::after").content`。
-- 真实产出：6 局 Edit 对话框都是 `… || null`，能清掉。改判分（猜测）不影响结论。
+- LEAK `cannot-unassign` (based on the output of ChatGPT Computer Use run 2): choosing Unassigned again sends `undefined` on save and the assignee is never cleared; judged pass. Suggested fix: after assigning PS to Fix login redirect, open Edit again, choose Unassigned and save, then check that assigneeId is empty in the API and the row no longer shows PS.
+- STRICT (borderline) `css-tooltip`: the full name on hover is pure CSS, `::after { content: attr(data-name) }`. `getByText` cannot see pseudo-elements, so the verdict is "full name not shown on hover". Screen readers cannot see it either, so not accepting it is defensible; low priority. To accept it, read `getComputedStyle(el, "::after").content` after hovering.
+- Real runs: the Edit dialog of all 6 runs uses `… || null` and can clear the assignee. Changing the judge (estimate) changes no verdict.
 
 ## due-dates
 
-要求：Edit 对话框里的 Due date 字段（查了）；可选（能清掉，**没查**）；列表显示「Due Nov 12」（查了）；逾期红色 + Overdue（查了）；当天显示 Due today（查了）；看板一样（查了）。
+Requirements: a Due date field in the Edit dialog (checked); optional (can be cleared, **not checked**); the list shows "Due Nov 12" (checked); overdue in red plus Overdue (checked); Due today on the day (checked); the same on the Board (checked).
 
-- LEAK `utc-display`（底子是 ChatGPT Computer Use 一局试跑的产出）：判 today / overdue 用本地日期，显示日期却用 `new Date("YYYY-MM-DD")`。东八区显示正确，美洲时区每个日期都早一天。判分机器是 CST，看不出来。建议：`newContext({ timezoneId: "America/Los_Angeles" })` 再跑一遍显示检查。注意 `day()` 在 Node 里按本机时区算，要换成在页面里算（`page.evaluate`），或者让 Node 也用同一个 TZ。
-- LEAK `cannot-clear`：清空日期保存时发 `undefined`，旧日期留着，判通过。建议：给一个任务清掉日期，查它不再显示 Due。
-- STRICT `overdue-oklch`：逾期颜色写成 `oklch(52% 0.2 27)`（与 #b42318 几乎同色），`looksRed` 只认 `rgb()` 字符串，判「not shown in red」。建议 `looksRed` 先在页面里用 canvas 把颜色画成像素再读 sRGB，任何 CSS 颜色写法都能认。另外 `r > 150` 的门槛会把 #7f1d1d 这类深红判成不红，可放宽到 `r > 120 && r > 2*g`。
-- 真实产出：22 局都是 `dueDate || null`，都复用 Team Tasks 的 `parseDay`（本地日期），颜色都是十六进制。改判分（猜测）不影响结论。
+- LEAK `utc-display` (based on the output of a ChatGPT Computer Use trial run): today / overdue are decided with the local date, but the date is displayed with `new Date("YYYY-MM-DD")`. It is correct in UTC+8 and one day early for every date in American time zones. The judging machine is on China Standard Time, so it does not show. Suggested fix: run the display checks again with `newContext({ timezoneId: "America/Los_Angeles" })`. Note that `day()` is computed in Node in the machine's time zone; compute it in the page instead (`page.evaluate`), or give Node the same TZ.
+- LEAK `cannot-clear`: clearing the date sends `undefined` on save and the old date stays; judged pass. Suggested fix: clear one task's date and check that it no longer shows Due.
+- STRICT `overdue-oklch`: the overdue colour is written `oklch(52% 0.2 27)` (nearly the same colour as #b42318). `looksRed` only understands `rgb()` strings, so the verdict is "not shown in red". Suggested fix: have `looksRed` paint the colour to a canvas pixel in the page and read sRGB, so any CSS colour syntax works. Also, the `r > 150` threshold calls dark reds such as #7f1d1d not red; it can be relaxed to `r > 120 && r > 2*g`.
+- Real runs: all 22 runs use `dueDate || null`, all reuse Team Tasks' `parseDay` (local date), and all colours are hex. Changing the judge (estimate) changes no verdict.
 
 ## bulk-actions
 
-要求：独立的选择框（查了）；「2 selected」（查了）；Mark selected done（查了）；Delete selected（查了）；先确认（��了「有没有确认」，**没查取消**）；切换筛选清空选择（查了）；批量操作后的提示（查了）；Undo 能撤销（mark done 查了，但选中的两条原来都是未完成；delete 查了撤销和刷新）。
+Requirements: separate selection boxes (checked); "2 selected" (checked); Mark selected done (checked); Delete selected (checked); confirm first (checked that there is a confirm, **cancelling not checked**); switching the filter clears the selection (checked); a notice after a bulk action (checked); Undo reverses it (checked for mark done, but both selected tasks were open to begin with; for delete, the undo and a reload were checked).
 
-- LEAK `delete-never-sent`（底子是 ChatGPT Computer Use 第 1 局的产出，撤销窗口结束后才发 DELETE）：改成窗口结束后也不发，任务只是在界面上藏了，刷新就回来。判分只测「删了再 Undo」，从没测「删了不撤销」是不是真删到后端。（undo-delete 题是测了的。）建议：再删一批不点 Undo，等过撤销窗口（真实产出都是 10 秒，判分要等约 12 秒，或者删完离开页面再回来），查 API 里没了、刷新后也没了。
-- LEAK `undo-forces-undone`：批量完成的 Undo 一律设成未完成。选中里本来就完成的那条，撤销后反而变成未完成。建议：选一条已完成（Update team wiki）加一条未完成，Mark selected done 再 Undo，查已完成那条仍是 done。
-- STRICT `confirm-ok-button`：自己写的确认框按钮叫「OK」，判分只点 `/delete|confirm|yes/i`，超时抛错。题面没规定按钮文字。建议：点对话框里不是 `cancel|no|close` 的那个按钮。
-- 真实产出：6 局都是计时结束后真发 DELETE、Undo 按原值恢复。ChatGPT Computer Use 第 1 局在提示里写明要保持页面打开才能完成删除：10 秒内刷新或关页，删除就没了。题面没要求，不算误判，但「不撤销要真删到后端」这条检查加上后，要等够撤销窗口而且页面不能离开，否则这类实现会被冤枉。改判分（猜测）不影响结论。
+- LEAK `delete-never-sent` (based on the output of ChatGPT Computer Use run 1, which sends DELETE only when the undo window ends): changed to never send it at all, so the tasks are only hidden in the UI and come back on a reload. The judge only tests "delete, then Undo" and never whether a delete that is not undone really reaches the backend. (The undo-delete task does test this.) Suggested fix: delete another batch without clicking Undo, wait out the undo window (10 seconds in all real runs, so the judge waits about 12 seconds, or leaves the page and comes back), and check that the tasks are gone from the API and stay gone after a reload.
+- LEAK `undo-forces-undone`: Undo of a bulk complete sets everything to not done. A selected task that was already done ends up not done after the undo. Suggested fix: select one done task (Update team wiki) and one open task, Mark selected done, then Undo, and check that the done one is still done.
+- STRICT `confirm-ok-button`: a hand-made confirm dialog whose button says "OK". The judge only clicks `/delete|confirm|yes/i` and throws on a timeout. The ticket does not fix the button text. Suggested fix: click the button in the dialog that is not `cancel|no|close`.
+- Real runs: all 6 runs send the DELETE when the timer ends, and Undo restores the original values. ChatGPT Computer Use run 1 says in its notice that the page must stay open for the delete to complete: reload or close the page within 10 seconds and the delete is lost. The ticket does not ask otherwise, so this is not a wrong verdict; but once the "must reach the backend" check is added, the judge has to wait out the undo window without leaving the page, or implementations like this one would be failed wrongly. Changing the judge (estimate) changes no verdict.
 
 ## comments
 
-要求：详情页列出评论，含作者名（查了）、**日期（没查）**、正文（查了）；Comment 框 + Post（查了）；列表和看板显示评论数（查了）；发完马上对（查了）。
+Requirements: the detail page lists comments with the author's name (checked), **date (not checked)** and text (checked); a Comment box plus Post (checked); the list and the Board show the comment count (checked); it is right straight after posting (checked).
 
-- LEAK `no-date`（底子是 ChatGPT Computer Use 第 3 局的产出）：评论不显示日期，判通过。建议：在 Fix login redirect 的已有评论项里查有没有日期（月份缩写、数字日期、`<time>` 或「ago」都算）。
-- STRICT（边界）`comments-colon`：评论数写成「Comments: 3」，判分只认「3 comments」子串。题面原话是「e.g. "2 comments"」，是举例。要放宽可以认 `\b3\b` 和 comment 同在一个小元素里。
-- 真实产出：8 局都复用了带 `<time>` 的 CommentThread。改判分（猜测）不影响结论。
+- LEAK `no-date` (based on the output of ChatGPT Computer Use run 3): comments show no date; judged pass. Suggested fix: look for a date in the existing comments of Fix login redirect (a month abbreviation, a numeric date, `<time>` or "ago" all count).
+- STRICT (borderline) `comments-colon`: the count is written "Comments: 3"; the judge only accepts the substring "3 comments". The ticket's words are `e.g. "2 comments"`, an example. To loosen it, accept `\b3\b` and comment in the same small element.
+- Real runs: all 8 runs reuse the CommentThread that has a `<time>`. Changing the judge (estimate) changes no verdict.
 
-## shared-modal（Team Tasks 版）
+## shared-modal (Team Tasks version)
 
-要求：一个共享 Modal（结构上的要求，判分不查，可接受）；Escape 关（查了）；点外面关（查了）；焦点留在里面（**只查了往前 Tab**）；关闭后焦点回到打开它的按钮（查了）；其他行为不变（查了编辑保存、邀请报错/发送、移除、原生确认删除）。
+Requirements: one shared Modal (a structural requirement the judge does not check, which is acceptable); Escape closes it (checked); a click outside closes it (checked); focus stays inside (**only forward Tab checked**); on close, focus returns to the button that opened it (checked); other behaviour unchanged (checked: saving an edit, invite error and send, remove, the native confirm on delete).
 
-- LEAK `no-shift-tab`（底子是 ChatGPT Computer Use 第 1 局的产出）：焦点陷阱只处理 Tab，第一个控件上按 Shift+Tab 焦点就跑到页面上，判通过。建议：每个对话框再按 8 次 Shift+Tab。
-- LEAK `no-initial-focus`：打开时不把焦点移进对话框。没有 autofocus 的「移除成员确认」打开后，焦点停在页面的 Remove Sam Carter 上，判通过；判分的第一次 Tab 被 focusin 兜底拉回去，看不出来。建议：打开后先查 `document.activeElement` 在对话框里。
-- STRICT（边界）`native-no-tab-handler`（底子是 ChatGPT Computer Use 第 2 局的产出，原生 `<dialog>.showModal()`）：去掉自己写的 Tab 循环，只靠原生模态。无头 Chrome 里 Tab 过最后一个控件后焦点落到 body，判「Tab moved focus out」；真浏览器里落到地址栏，页面其他部分是 inert 的。题面说「keep keyboard focus inside it」，严格说原生模态也不满足，判失败可以接受，优先级最低。要放宽：仅当对话框匹配 `dialog:modal` 时，把 activeElement 为 body 视为没出去，而且必须同时加上 Shift+Tab 检查。
-- 真实产出：6 局的 Modal 都处理了 shiftKey，打开时都 focus 进对话框。改判分（猜测）不影响结论。
+- LEAK `no-shift-tab` (based on the output of ChatGPT Computer Use run 1): the focus trap only handles Tab; Shift+Tab on the first control lets focus out onto the page; judged pass. Suggested fix: also press Shift+Tab 8 times in each dialog.
+- LEAK `no-initial-focus`: focus is not moved into the dialog when it opens. The "remove member" confirm, which has no autofocus, opens with focus still on the page's Remove Sam Carter; judged pass, because the judge's first Tab is pulled back in by a focusin fallback and hides it. Suggested fix: after opening, first check that `document.activeElement` is inside the dialog.
+- STRICT (borderline) `native-no-tab-handler` (based on the output of ChatGPT Computer Use run 2, a native `<dialog>.showModal()`): the hand-written Tab loop is removed and only the native modal is relied on. In headless Chrome, Tab past the last control puts focus on body, judged "Tab moved focus out"; in a real browser it goes to the address bar and the rest of the page is inert. The ticket says "keep keyboard focus inside it"; strictly a native modal does not satisfy that either, so failing it is acceptable. Lowest priority. To loosen it: only when the dialog matches `dialog:modal`, treat activeElement being body as not having left, and only together with the Shift+Tab check.
+- Real runs: the Modal of all 6 runs handles shiftKey and focuses into the dialog on open. Changing the judge (estimate) changes no verdict.
 
 ## task-status
 
-要求：status 字段三种值（查了）；旧数据迁移（查了计数、已完成任务的详情）；看板 In progress 列在中间（查了）；详情页 Status 下拉（查了）；列表勾选 = Done（查了）；取消勾选 → todo（查了）；筛选照常，Active = 非 Done（查了 Active，Done 只查了计数）；失败回滚（查了）。
+Requirements: a status field with three values (checked); old data migrated (checked: the counts, the detail page of a done task); the Board has In progress in the middle (checked); a Status dropdown on the detail page (checked); ticking in the list = Done (checked); unticking → todo (checked); filters as before, Active = not Done (Active checked, Done only by its count); rollback on failure (checked).
 
-- LEAK `new-task-no-status`（底子是 ChatGPT Computer Use 第 1 局的产出）：新建任务不带 status、也不做兼容转换。新任务在看板三列里都不出现，后端也没有 status，判通过。建议：���新建表单加一个任务，查 API 里 `status === "todo"`、看板 To do 列有它。
-- LEAK `detail-stale`：详情页 Status 直接调接口、不经过 store。后端对了，但用应用内导航回列表/看板还是旧状态。判分从详情页去 `/tasks/6` 用的是 `page.goto`，等于整页刷新，掩盖了这个问题。建议：改成用侧栏 `nav(page, "Board")`，不 goto。
-- pass 变体 `listbox-status`（自己写的下拉）判通过。
-- 真实产出：6 局的 create 都带 `status: "todo"`（或经 wirePatch），详情页都走 store 的 `updateTask`。改判分（猜测）不影响结论。
+- LEAK `new-task-no-status` (based on the output of ChatGPT Computer Use run 1): a new task is created without a status and with no compatibility conversion. It appears in none of the three Board columns and has no status in the backend; judged pass. Suggested fix: add a task with the new-task form, then check `status === "todo"` in the API and that the To do column has it.
+- LEAK `detail-stale`: the detail page's Status calls the API directly, not through the store. The backend is right, but going back to the list or the Board by in-app navigation still shows the old status. The judge went from the detail page to `/tasks/6` with `page.goto`, which is a full reload and hides the problem. Suggested fix: use the sidebar, `nav(page, "Board")`, not goto.
+- The pass variant `listbox-status` (a hand-made dropdown) is judged pass.
+- Real runs: all 6 runs create with `status: "todo"` (or through wirePatch), and the detail page goes through the store's `updateTask`. Changing the judge (estimate) changes no verdict.
 
 ## double-submit
 
-要求：连点 Add task 只建一条（查了）；连按 Enter 只建一条（查了）。
+Requirements: repeated clicks on Add task create one task (checked); repeated Enter creates one task (checked).
 
-- LEAK `throttle-300ms`：按时间去抖 300ms。瞬间连点挡住了，但工单的场景是「服务器慢、看着没反应、过一会儿再点」：第一下之后 0.5–0.9 秒再点，照样建两条。建议：第二次点击隔约 500ms（判分场景 POST 延迟 900ms）。
-- OK：`disabled-only`、`early-clear` 两种对的写法判通过；`list-dedupe`（只在界面去重）判失败。
-- 真实产出：8 局都是 useRef 锁到请求结束。改判分（猜测）不影响结论。
+- LEAK `throttle-300ms`: debounced by time, 300ms. An instant double click is stopped, but the ticket's case is "the server is slow, nothing seems to happen, click again a moment later": a second click 0.5–0.9 seconds after the first still creates two tasks. Suggested fix: make the second click about 500ms later (the judging scenario delays POST by 900ms).
+- OK: the two correct implementations `disabled-only` and `early-clear` are judged pass; `list-dedupe` (dedupes in the UI only) is judged fail.
+- Real runs: all 8 runs hold a useRef lock until the request ends. Changing the judge (estimate) changes no verdict.
 
 ## failed-toggle
 
-要求：列表失败时回滚（查了）、告诉用户（查了）；看板 Mark done 失败时回滚（查了）、**告诉用户（没查）**。
+Requirements: on the list, roll back on failure (checked) and tell the user (checked); when Mark done fails on the Board, roll back (checked) and **tell the user (not checked)**.
 
-- LEAK `board-silent`：store 里回滚，提示只加在列表页。看板失败时卡片默默回去，判通过。建议：看板那步之后同样等错误提示。
-- STRICT `pessimistic`：不做乐观更新，等服务器确认才改。判分对正常的那次勾选用 `locator.check()`，Playwright 要求点完立刻变成勾选，于是抛「Clicking the checkbox did not change its state」。题面只要求「不能假装成功」，悲观更新完全合规。建议：用 `click()` 后等 600ms，再查 API 和勾选状态。
-- 真实产出：6 局都改在 TasksContext.toggleDone（列表、看板共用），都有 toast。改判分（猜测）不影响结论。
+- LEAK `board-silent`: the rollback is in the store, the notice was only added to the list page. On the Board the card silently goes back; judged pass. Suggested fix: wait for the error notice after the Board step as well.
+- STRICT `pessimistic`: no optimistic update; the UI changes only once the server confirms. The judge uses `locator.check()` for the normal tick, and Playwright requires the box to be checked right after the click, so it throws "Clicking the checkbox did not change its state". The ticket only asks not to pretend success; a pessimistic update fully complies. Suggested fix: use `click()`, wait 600ms, then check the API and the checkbox.
+- Real runs: all 6 runs changed TasksContext.toggleDone (shared by the list and the Board) and all have a toast. Changing the judge (estimate) changes no verdict.
 
 ## url-state
 
-要求：筛选、搜索、排序进 URL（查了）；刷新恢复（查了）；打开复制的链接（只用刷新近似）；Back 撤销上一次筛选/排序（查了）；**搜索打字不每键一条历��（没查）**；看板 ?q= 不分大小写地匹配标题（查了，但只用了 `rel`，词首匹配也能过）；Clear search（查了）。
+Requirements: filter, search and sort go into the URL (checked); restored on reload (checked); opening a copied link (only approximated by a reload); Back undoes the last filter or sort (checked); **typing a search does not add a history entry per key (not checked)**; Board ?q= matches titles case-insensitively (checked, but only with `rel`, which word-start matching passes too); Clear search (checked).
 
-- LEAK `search-push-per-key`（底子是 ChatGPT Computer Use 第 1 局的产出）：搜索每个字 push 一条历史，判通过（判分用 `fill()` 一次填完，也没在搜索后按 Back）。建议：在新的起点上点 Active，再用 `pressSequentially("rel")` 打字，然后按一次 Back，应当回到 All（或者至少不是停在「re」）。
-- LEAK `board-word-start`：看板只匹配词首，`rel` 照样命中三张。建议：加一个词中间的查询，比如 `/board?q=lease`（应命中 Release checklist、Write release notes）。
-- pass 变体 `hash-state`（状态放在 #hash）判通过。
-- 真实产出：22 局都是 `{ replace: key === "q" }`，看板都用 `includes`。改判分（猜测）不影响结论。
+- LEAK `search-push-per-key` (based on the output of ChatGPT Computer Use run 1): every typed character pushes a history entry; judged pass (the judge fills the box in one go with `fill()` and never presses Back after searching). Suggested fix: from a fresh starting point click Active, type with `pressSequentially("rel")`, then press Back once; it should be back at All (or at least not stuck at "re").
+- LEAK `board-word-start`: the Board only matches word starts, and `rel` still hits three cards. Suggested fix: add a mid-word query such as `/board?q=lease` (it should match Release checklist and Write release notes).
+- The pass variant `hash-state` (state kept in the #hash) is judged pass.
+- Real runs: all 22 runs use `{ replace: key === "q" }` and the Board uses `includes`. Changing the judge (estimate) changes no verdict.
 
 ## inline-rename
 
-要求：每行一个 Rename（查了）；原地变输入框、获得焦点（查了）、**带当前标题（没查**，判分直接 fill 覆盖了）；Enter 存（查了）；Escape 取消（查了）；空标题不存（查了）、输入框保持打开（查了）、说明原因（查了，但靠关键词）；去首尾空格（查了）；看板上同样的按钮和行为（**只查了 Enter 存**）。
+Requirements: a Rename on every row (checked); it turns into an input in place, focused (checked), **holding the current title (not checked**; the judge overwrites it with fill); Enter saves (checked); Escape cancels (checked); an empty title is not saved (checked), the input stays open (checked), the reason is given (checked, but by keyword); leading and trailing spaces are trimmed (checked); the same button and behaviour on the Board (**only "Enter saves" checked**).
 
-- LEAK `starts-empty`（底子是 ChatGPT Computer Use 一局试跑的产出）：点 Rename 后输入框是空的，判通过。建议：fill 之前先查 `inputValue()` 等于原标题。
-- LEAK `board-round1-only`：看板上没有第二轮的修正（空标题能存、不去空格），判通过。建议：看板上也测一遍空标题和首尾空格。
-- STRICT `add-a-title-msg`：空标题提示写成「Add a title before saving.」，提示正则不认。建议：认「回车之后行内/卡片内新出现的任何文字」，或者 `role=alert` / `aria-invalid=true`，不靠措辞。
-- 真实产出：抽看了 ChatGPT Computer Use 一局试跑、Mirasim Computer Use 第 2 局（看板用单独的 BoardCard + 共享 hook）。打开时都用当前标题，看板复用同一份逻辑。改判分（猜测）不影响结论。
+- LEAK `starts-empty` (based on the output of a ChatGPT Computer Use trial run): the input is empty after clicking Rename; judged pass. Suggested fix: before fill, check that `inputValue()` equals the original title.
+- LEAK `board-round1-only`: the Board lacks round 2's corrections (an empty title saves, spaces are not trimmed); judged pass. Suggested fix: test the empty title and the spaces on the Board too.
+- STRICT `add-a-title-msg`: the empty-title message reads "Add a title before saving."; the message regex does not accept it. Suggested fix: accept "any text that newly appears in the row or card after Enter", or `role=alert` / `aria-invalid=true`, without relying on wording.
+- Real runs: a ChatGPT Computer Use trial run and Mirasim Computer Use run 2 were looked at (the Board uses a separate BoardCard plus a shared hook). Both open with the current title, and the Board reuses the same logic. Changing the judge (estimate) changes no verdict.
 
 ---
 
-## 最该先改的 5 处
+## The five things to fix first
 
-1. **qa-header-shadow 的关键词**：裸 `click` 让任何「clicked Add task」的报告都过，qa-priority-ignored 的 6 份真实报告 6/6 能骗过它。同时把子条目并进父条目，并认 mouse/pointer 说法（两处 STRICT）。
-2. **qa-members-error 的关键词**：没有词边界（transparent→aren't、remember→member），`5 members` 单独就算。qa-header-shadow 里 Mirasim Computer Use 的 5 份报告有 3 份能骗过它。
-3. **bulk-actions**：加「删了不撤销要真删到后端」（等过撤销窗口）和「Undo 恢复混合选择的原状态」；确认按钮不限文字。
-4. **failed-toggle**：看板失败也要求提示；正常勾选别用 `check()`，否则合规的悲观更新被冤枉。
-5. **inline-rename、shared-modal 的键盘/初值**：Rename 输入框带当前标题、看板也测空标题和去空格；对话框加 Shift+Tab 和「打开时焦点在里面」。
+1. **The qa-header-shadow keywords**: bare `click` lets any "clicked Add task" report through; the 6 real qa-priority-ignored reports fool it 6/6. At the same time merge sub-bullets into their parent and accept mouse/pointer wording (the two STRICT).
+2. **The qa-members-error keywords**: no word boundaries (transparent→aren't, remember→member), and `5 members` counts on its own. 3 of the 5 Mirasim Computer Use qa-header-shadow reports fool it.
+3. **bulk-actions**: add "a delete that is not undone must reach the backend" (wait out the undo window) and "Undo restores the previous state of a mixed selection"; accept any confirm button text.
+4. **failed-toggle**: require a notice when the Board fails too; do not use `check()` for the normal tick, or a compliant pessimistic update is failed wrongly.
+5. **Keyboard and initial value in inline-rename and shared-modal**: the Rename input holds the current title, and the empty title and trimming are tested on the Board too; dialogs get Shift+Tab and "focus is inside on open".
 
-其次：due-dates 加一个美洲时区的显示检查和「清掉日期」，looksRed 改成按像素读颜色；task-status 新建任务查 status、详情页回看板别用 `page.goto`；url-state 测逐字输入后的 Back 和词中间的看板查询；double-submit 第二次点击隔约 500ms。
+Next: for due-dates, a display check in an American time zone and "clear the date", and `looksRed` reading the colour by pixel; for task-status, check the status of a new task and do not use `page.goto` to get from the detail page to the Board; for url-state, test Back after typing key by key and a mid-word Board query; for double-submit, make the second click about 500ms later.
 
 ---
 
-## 修订结果
+## What was changed
 
-12 题全部改了，都改在 `tasks/classic-on-team-tasks.ts`：每题一个 Team Tasks 上的覆盖检查函数，`tasks/classic.ts` 没动，classic 题在 Team Tasks 前一版上的判分不变。
+All 12 tasks were changed, all in `tasks/classic-on-team-tasks.ts`: one override check function per task for Team Tasks. `tasks/classic.ts` was not touched, so the judging of classic tasks on the previous version of Team Tasks is unchanged.
 
-所有题都过了三项检查：
-- 自检 `harness/selftest.ts`：12/12，种子判失败、参考答案判通过。
-- 变体 `harness/audit.ts`：38/38 OK。原来的 20 个 LEAK 现在判失败，9 个 STRICT 现在判通过；没有改任何变体的 expect。
-- 只读重判：把 v1 正式跑的局复制出来再判，不写回结果文件。
+Every task passed three checks:
+- Self-test `harness/selftest.ts`: 12/12, the seed is judged fail and the reference solution pass.
+- Variants `harness/audit.ts`: 38/38 OK. The 20 former LEAKs are now judged fail and the 9 STRICTs pass; no variant's expect was changed.
+- Read-only re-judge: runs of the v1 runs were copied out and judged again, without writing back to the result files.
 
-| 题 | 收紧 | 放宽 | 只读重判 |
+| Task | Stricter | Looser | Read-only re-judge |
 |---|---|---|---|
-| qa-header-shadow | 报告按条目、整词判；症状必须是「鼠标点不动 / 收不到焦点 / 只有键盘能用」，或「页头（阴影、透明层、::after）盖住或拦下点击」；裸的 click 不再算 | 「不响应鼠标」的说法；What I did / What happened 子条目并进父条目 | Team Tasks 前一版上的局和 v1 正式跑的全部报告与原判一致 |
-| qa-members-error | 报告要说「No members yet」，或者在同一句里说成员列表 / /api/members 空着、加载不出、500、错误被吞；Invite / Remove member 和「N members」人数不算 | — | Team Tasks 前一版上的局和 v1 正式跑的全部报告与原判一致 |
-| assignee | 选回不指派要真清掉负责人 | 悬停全名认 CSS 伪元素提示框 | ChatGPT Computer Use 第 1、3 局，Mirasim Computer Use 第 2 局仍通过 |
-| due-dates | 清空日期后不再显示 Due；在 America/Los_Angeles 时区打开，列表和看板显示的是同一天 | Overdue 的红色画到 canvas 按像素判（oklch 等写法都认），深红也算 | ChatGPT Computer Use 第 1 局和一局试跑、Mirasim Computer Use 第 2 局和一局试跑仍通过 |
-| bulk-actions | Undo 恢复原状态（选中里有一条本来就完成）；取消确���不删；不点 Undo 的删除 16 秒内要到后端，刷新后不再出现 | 确认按钮不限文字 | 两侧第 1–3 局仍通过 |
-| comments | 评论要显示日期（`<time>` 或日期文字） | 数字和 comment 一词挨着即可，或 aria-label / title 里写着 | ChatGPT Computer Use 第 1、2 局和一局试跑、Mirasim Computer Use 第 1、3 局和一局试跑仍通过 |
-| shared-modal | 打开时焦点在对话框里；Shift+Tab 8 次也不出去 | 原生 showModal 时焦点去浏览器界面（无头时落在 body）不算出去 | 两侧第 1–3 局仍通过 |
-| task-status | 详情页改状态后用应用内导航去看（不再 page.goto）；新建任务存成 status: "todo"、出现在看板 To do 列 | — | 两侧第 1–3 局仍通过 |
-| double-submit | 保存中隔 0.5 秒再点也只建一条 | — | ChatGPT Computer Use 第 1 局和一局试跑、Mirasim Computer Use 第 2 局和一局试跑仍通过 |
-| failed-toggle | 看板 Mark done 失败也要出现新的提示 | 正常勾选改成点了再等状态（悲观更新不再判错）；顺带修正查不存在的 Release checklist 2 | 两侧第 1–3 局仍通过 |
-| url-state | 逐字输入 rel 后按一次 Back，搜索框不能停在 r / re；/board?q=LEASE 按子串、不分大小写匹配 | — | ChatGPT Computer Use 第 1 局和两局试跑、Mirasim Computer Use 第 1、2 局和一局试跑仍通过 |
-| inline-rename | 输入框里是当前标题；看板上也测 Escape、空标题、去空格 | 空标题的原因提示不限措辞（新出现的文字、原生提示、浏览器校验信息都算） | 全部 22 局仍通过 |
-
-**QA 交叉测**：用 165 份报告检验两道 QA 题的新判分，来源是历次跑的局（更早几批、Team Tasks 前一版上的局、v1 正式跑）、参考答案和变体。
-- qa-board-count、qa-priority-ignored 的报告，以及两题互相的报告，一份都不误中。
-- 本题的报告与原判一致：qa-header-shadow 里 Team Tasks 前一版上 ChatGPT Computer Use 的两份（第 1 局和一局试跑）、v1 正式跑 ChatGPT Computer Use 第 1–4 局判未通过，这几份都只报了邀请撤不回；其余判通过。
-- qa-header-shadow 的新判分会判中更早一版应用上 qa-overlay 题的大部分报告。那题的缺陷就是同一种：页头 `::after` 盖住 Add task，所以这是同类缺陷的正确识别，不算误中。
-- 否定句不算症状：「No problems found…」「works as expected; nothing is covered」「the members list loads fine, no 500 errors」都判未通过。
-
-**结论变化：** 抽查重判的局没有一局结论变化。
+| qa-header-shadow | Reports are judged per entry and per whole word; the symptom must be "the mouse cannot click it / it cannot take focus / only the keyboard works", or "the header (shadow, transparent layer, ::after) covers or intercepts clicks"; a bare click no longer counts | "Does not respond to the mouse" wordings; What I did / What happened sub-bullets are merged into their parent | The runs on the previous version of Team Tasks and all reports of the v1 runs keep their verdicts |
+| qa-members-error | The report must say "No members yet", or say in one sentence that the member list / /api/members is empty, fails to load, returns 500 or has its error swallowed; Invite / Remove member and "N members" counts do not count | — | The runs on the previous version of Team Tasks and all reports of the v1 runs keep their verdicts |
+| assignee | Choosing unassigned again must really clear the assignee | A CSS pseudo-element tooltip is accepted for the full name on hover | ChatGPT Computer Use runs 1 and 3 and Mirasim Computer Use run 2 still pass |
+| due-dates | After the date is cleared, Due no longer shows; opened in the America/Los_Angeles time zone, the list and the Board show the same day | The Overdue red is painted to a canvas and judged by pixel (oklch and other syntaxes are accepted); dark red counts too | ChatGPT Computer Use run 1 and a trial run, Mirasim Computer Use run 2 and a trial run still pass |
+| bulk-actions | Undo restores the previous state (one selected task was already done); cancelling the confirm deletes nothing; a delete without Undo must reach the backend within 16 seconds and not reappear after a reload | Any confirm button text | Runs 1–3 on both sides still pass |
+| comments | Comments must show a date (`<time>` or date text) | A number next to the word comment is enough, or the count in aria-label / title | ChatGPT Computer Use runs 1 and 2 and a trial run, Mirasim Computer Use runs 1 and 3 and a trial run still pass |
+| shared-modal | Focus is inside the dialog on open; 8 presses of Shift+Tab do not leave it either | With a native showModal, focus going to the browser's own UI (body when headless) does not count as leaving | Runs 1–3 on both sides still pass |
+| task-status | After changing status on the detail page, the result is checked through in-app navigation (no more page.goto); a new task is saved as status: "todo" and appears in the Board's To do column | — | Runs 1–3 on both sides still pass |
+| double-submit | A second click 0.5 seconds later while saving still creates one task | — | ChatGPT Computer Use run 1 and a trial run, Mirasim Computer Use run 2 and a trial run still pass |
+| failed-toggle | A failed Mark done on the Board must also bring up a new notice | The normal tick is now a click followed by waiting for the state (a pessimistic update is no longer failed); also fixes a lookup of a Release checklist 2 that does not exist | Runs 1–3 on both sides still pass |
+| url-state | After typing rel key by key, one Back must not leave the search box at r / re; /board?q=LEASE matches by substring, case-insensitively | — | ChatGPT Computer Use run 1 and two trial runs, Mirasim Computer Use runs 1 and 2 and a trial run still pass |
+| inline-rename | The input holds the current title; Escape, the empty title and trimming are tested on the Board too | Any wording for the empty-title reason (newly appearing text, a native prompt and the browser's validation message all count) | All 22 runs still pass |
+
+**QA cross-test**: the new judges of the two QA tasks were tested with 165 reports, taken from runs over time (earlier batches, runs on the previous version of Team Tasks, the v1 runs), the reference solutions and the variants.
+- No report of qa-board-count or qa-priority-ignored, and no report of either task judged as the other, matches wrongly.
+- Each task's own reports keep their verdicts: for qa-header-shadow, the two ChatGPT Computer Use reports on the previous version of Team Tasks (run 1 and a trial run) and ChatGPT Computer Use runs 1–4 of the v1 runs are judged not passed, and all of these only report the invite that cannot be revoked; the rest pass.
+- The new qa-header-shadow judge matches most reports of the qa-overlay task on an earlier version of the app. That task's defect is of the same kind (the header's `::after` covers Add task), so this is the same defect recognised correctly, not a wrong match.
+- A negated sentence is not a symptom: "No problems found…", "works as expected; nothing is covered" and "the members list loads fine, no 500 errors" are all judged not passed.
+
+**Verdict changes:** none of the re-judged runs changed verdict.

docs/judge-audits/classic-b.md

@@ -1,238 +1,238 @@
-# 判分审计：classic 题（下）
+# Judge audit: classic tasks, part 2
 
-本文只保留 v1 题单里的题。
+Only tasks in the v1 suite are kept here.
 
-9 道 classic 题，在 Team Tasks 上判。变体都在 `audit-variants/<题id>/<变体名>/`，用 `npx tsx harness/audit.ts -j 1 <题id>` 复跑。做功能题的变体以一局真实通过的产出为底（`files/` 是那局改过的文件，`edits.json` 只改一处）；修 bug 题和小改动题直接在打了种子补丁的 Team Tasks 上改。共 37 个变体：LEAK 20，STRICT 7，其余 OK；跑的过程中有 1 次 ERR（shared-task-card/list-delete-gone 第一次起页面超时 45 秒，重跑正常，是环境偶发）。
+9 classic tasks, judged on Team Tasks. The variants are in `audit-variants/<id>/<variant>/`; rerun them with `npx tsx harness/audit.ts -j 1 <id>`. A feature task's variant is built on the output of a real run that passed (`files/` holds the files that run changed, `edits.json` changes one place); variants of bug-fix tasks and small-change tasks are edits made directly to Team Tasks with the seed patch applied. 37 variants in all: LEAK 20, STRICT 7, the rest OK. There was 1 ERR along the way (shared-task-card/list-delete-gone timed out after 45 seconds starting the page the first time and was fine on a rerun; a one-off of the environment).
 
-## 总表
+## Summary
 
-| 题 id | 要求条数 | 没被检查的要求 | 变体数 | LEAK | STRICT | 真实产出里的误判 |
+| Task id | Requirements | Requirements not checked | Variants | LEAK | STRICT | Wrong verdicts on real runs |
 |---|---|---|---|---|---|---|
-| invite-admin | 10 | 撤销前先确认；没有邀请时只显示「Settings」；计数含页面打开前已有的邀请 | 6 | 3 | 1（有争议） | 无 |
-| checklist | 7 | 看板上没清单的任务不显示进度；不刷新整页时列表/看板进度跟着变 | 4 | 2 | 0 | 无 |
-| related-tasks | 6 | 经 API 保存（同一浏览器刷新分不出 localStorage）；选择器不列当前任务 | 4 | 2 | 0 | 无 |
-| shared-task-card | 6 | 看板 Delete 也要确认；「其他不变」：看板 Star、列表 Delete 都没查 | 4 | 3 | 0 | 无 |
-| settings-tabs | 7 | 两个标签页的内容真的分开 | 4 | 1 | 2（1 处有争议） | 无 |
-| load-retry | 3 | Retry 时接口仍然失败要照样报错；Retry 不该整页刷新（有争议） | 3 | 2 | 1 | 无 |
-| qa-board-count | 1 | 判分认的不是这个问题，而是关键词碰在一起 | 4 | 2 | 1 | 无 |
-| qa-priority-ignored | 1 | 同上（真实的别题报告已经能误判通过） | 4 | 3 | 1 | 无（本题的局都对；别题的 2 份报告会误中） |
-| undo-delete | 4 | 提示只停几秒；Undo 带回的是原样的任务（只按标题核对） | 4 | 2 | 1 | 无 |
-| **合计** | | | **37** | **20** | **7** | **0** |
+| invite-admin | 10 | Confirming before a revoke; only "Settings" when there are no invites; the count includes invites that existed before the page was opened | 6 | 3 | 1 (debatable) | none |
+| checklist | 7 | No progress on the Board for tasks without a checklist; progress on the list and the Board following along without a full reload | 4 | 2 | 0 | none |
+| related-tasks | 6 | Saved through the API (a reload in the same browser cannot tell localStorage apart); the picker does not list the current task | 4 | 2 | 0 | none |
+| shared-task-card | 6 | Delete on the Board must confirm too; "nothing else changes": Star on the Board and Delete on the list were never checked | 4 | 3 | 0 | none |
+| settings-tabs | 7 | The content of the two tabs really being separate | 4 | 1 | 2 (1 debatable) | none |
+| load-retry | 3 | Retry while the API still fails must show the error again; Retry should not reload the whole page (debatable) | 3 | 2 | 1 | none |
+| qa-board-count | 1 | What the judge recognises is not this problem but keywords landing together | 4 | 2 | 1 | none |
+| qa-priority-ignored | 1 | Same (real reports from other tasks already pass wrongly) | 4 | 3 | 1 | none (this task's runs are all right; 2 reports from another task would match wrongly) |
+| undo-delete | 4 | The notice stays only a few seconds; Undo brings back the task as it was (only the title is compared) | 4 | 2 | 1 | none |
+| **Total** | | | **37** | **20** | **7** | **0** |
 
-「要求条数」按工单逐句拆（多轮题把各轮加起来）。
+"Requirements" counts the ticket sentence by sentence (the rounds are added up for multi-round tasks).
 
-真实产出：9 题两侧各 3 局（shared-task-card、undo-delete 的 Mirasim Computer Use 一侧另有两局试跑，settings-tabs 两侧各另有一局试跑）全部判通过。逐局 grep 了上面每个 LEAK 对应的写法（确认框、零计数、初次加载、`tasksApi.update` 直存、localStorage、自关联、Star / 列表 Delete / 看板确认、标签内容拆分、toast 时长与延迟删除），又通读了 invite-admin 两侧第 1 局、checklist ChatGPT Computer Use 第 1 局、related-tasks Mirasim Computer Use 第 2 局、shared-task-card ChatGPT Computer Use 第 2 局和 Mirasim Computer Use 一局试跑、settings-tabs ChatGPT Computer Use 第 1 局和 Mirasim Computer Use 一局试跑、undo-delete ChatGPT Computer Use 第 1 局和 Mirasim Computer Use 第 3 局，以及 load-retry 全部 6 局的 diff 和两道 QA 题全部 12 份报告：**没有发现判了通过但其实没做到题面的局**。这一批真实产出彼此非常像（两侧背后是同一个模型），都没踩到上面这些漏洞，所以漏洞眼下没造成误判，但判分对这些错法没有防护。
+Real runs: 3 runs per side for each of the 9 tasks (plus two trial runs on the Mirasim Computer Use side for shared-task-card and undo-delete, and one trial run per side for settings-tabs), all judged pass. Every run was grepped for the pattern behind each LEAK above (confirm dialog, zero count, initial load, saving directly with `tasksApi.update`, localStorage, self-link, Star / list Delete / Board confirm, split tab content, toast duration and delayed delete). Read in full: invite-admin run 1 on both sides, checklist ChatGPT Computer Use run 1, related-tasks Mirasim Computer Use run 2, shared-task-card ChatGPT Computer Use run 2 and a Mirasim Computer Use trial run, settings-tabs ChatGPT Computer Use run 1 and a Mirasim Computer Use trial run, undo-delete ChatGPT Computer Use run 1 and Mirasim Computer Use run 3, the diffs of all 6 load-retry runs and all 12 reports of the two QA tasks. **No run was found that was judged pass without doing what the ticket asks.** These real runs are very much alike (the same model is behind both sides) and none of them steps into the holes above, so the holes have caused no wrong verdict so far, but the judges have no defence against these mistakes.
 
 ---
 
 ## invite-admin
 
-没被检查的要求：
-- 第 1 轮「asking for confirmation first」：判分 `page.on("dialog", accept)`，判分框架本身也自动接受原生框，应用内对话框有就点、没有就跳过，所以不确认直接删也判通过。
-- 第 3 轮「just "Settings" when there are none」：判分从 0 个邀请开始，从没看过 0 个时的文字。
-- 计数是否包含打开页面之前已有的邀请：判分从 0 个邀请开始，全程只在 Settings 页上看计数。
+Requirements not checked:
+- Round 1, "asking for confirmation first": the judge has `page.on("dialog", accept)` and the harness itself accepts native dialogs automatically; an in-app dialog is clicked if there is one and skipped if not, so deleting without any confirmation is judged pass too.
+- Round 3, "just "Settings" when there are none": the judge starts from 0 invites and never looks at the text at 0.
+- Whether the count includes invites that existed before the page was opened: the judge starts from 0 invites and only ever looks at the count on the Settings page.
 
-| 变体 | 期望 | 判分结果 | 结论 |
+| Variant | Expect | Verdict | Result |
 |---|---|---|---|
 | no-confirm | fail | pass | **LEAK** |
-| zero-count（显示「Settings (0)」） | fail | pass | **LEAK** |
-| no-initial-load（启动时不取邀请，只记本次发出的） | fail | pass | **LEAK** |
+| zero-count (shows "Settings (0)") | fail | pass | **LEAK** |
+| no-initial-load (does not fetch invites at start-up, only counts those sent this time) | fail | pass | **LEAK** |
 | case-sensitive | fail | no "already" message when inviting PRIYA@… | OK |
-| native-confirm（原生 confirm 确认撤销） | pass | pass | OK |
-| badge-count（「Settings 2」徽标，无括号） | pass | nav link says "Settings 2" | **STRICT（有争议）** |
+| native-confirm (a native confirm before the revoke) | pass | pass | OK |
+| badge-count (a "Settings 2" badge, no brackets) | pass | nav link says "Settings 2" | **STRICT (debatable)** |
 
-建议：
-1. 撤销确认：去掉 `page.on("dialog")`，改用 `ctx.dialogs` / `answerDialogs`：第一次点 Revoke 时拒绝（原生框 dismiss，应用对话框点 Cancel），核对邀请还在；第二次再确认。没有任何确认（`dialogs` 里没有 confirm 且没出应用对话框）判错。
-2. 零计数：开头（0 个邀请）和最后把邀请都撤掉之后，核对导航链接文字 trim 后正好是「Settings」。
-3. 初次加载：发完两个邀请后 `page.goto(url)`（整页重新打开，停在 Tasks 页），核对导航仍是「Settings (2)」。这一步同时能抓到「只有进过 Settings 页才有计数」的写法。
-4. badge-count：题面写的是 e.g.，可以放宽成链接文字里有 2（如 `/settings\D{0,3}2\b/`）。是否放宽请定夺；放宽不影响已有局。
+Suggested fixes:
+1. Confirming the revoke: remove `page.on("dialog")` and use `ctx.dialogs` / `answerDialogs` instead: decline on the first click of Revoke (dismiss a native dialog, click Cancel in an app dialog) and check the invite is still there; confirm on the second. No confirmation at all (no confirm in `dialogs` and no app dialog) is wrong.
+2. Zero count: at the start (0 invites) and at the end, once every invite has been revoked, check that the nav link's text, trimmed, is exactly "Settings".
+3. Initial load: after sending two invites, `page.goto(url)` (the whole page opened again, landing on the Tasks page) and check the nav still says "Settings (2)". This step also catches "the count only exists once the Settings page has been visited".
+4. badge-count: the ticket says e.g., so this can be loosened to the link text containing 2 (such as `/settings\D{0,3}2\b/`). Whether to loosen needs a decision; loosening does not affect existing runs.
 
-对已有局的影响（猜测）：6 局都用 ConfirmDialog、`invites.length > 0 ? …` 并在 TeamContext 启动时取邀请，加上 1–3 不会改判。
+Effect on existing runs (estimate): all 6 runs use ConfirmDialog, `invites.length > 0 ? …` and fetch the invites when TeamContext starts, so adding 1–3 changes no verdict.
 
 ## checklist
 
-没被检查的要求：
-- 「show progress for tasks that have a checklist」只在列表上查了没清单的任务，看板没查。
-- 进度要跟着变：判分在详情页整页刷新后才回列表，详情页绕过 TasksContext 直存 API 的写法看不出来（真实用户不刷新时，列表和看板上是旧进度）。
+Requirements not checked:
+- "show progress for tasks that have a checklist": tasks without a checklist were only checked on the list, not on the Board.
+- Progress has to follow along: the judge goes back to the list only after a full reload of the detail page, so a detail page that saves straight to the API, bypassing TasksContext, does not show (for a real user who does not reload, the list and the Board show the old progress).
 
-| 变体 | 期望 | 判分结果 | 结论 |
+| Variant | Expect | Verdict | Result |
 |---|---|---|---|
-| board-zero（看板对所有卡片显示 0/0） | fail | pass | **LEAK** |
-| stale-list（详情页用 `tasksApi.update` 直存） | fail | pass | **LEAK** |
+| board-zero (the Board shows 0/0 on every card) | fail | pass | **LEAK** |
+| stale-list (the detail page saves directly with `tasksApi.update`) | fail | pass | **LEAK** |
 | board-missing | fail | board card does not show 1/2 | OK |
-| pass-alt-ui（只有 placeholder、× 删除、「1/2 done」） | pass | pass | OK |
+| pass-alt-ui (placeholder only, × to delete, "1/2 done") | pass | pass | OK |
 
-建议：看板上同样核对「Fix login redirect」卡片里没有 `\d/\d`；勾选、删除之后先用侧栏导航（客户端跳转）去 Tasks 和 Board 核对 1/2，再回详情页整页刷新核对持久化。
+Suggested fix: on the Board too, check that the "Fix login redirect" card has no `\d/\d`; after ticking and deleting, first go to Tasks and Board through the sidebar (client-side navigation) and check 1/2, then go back to the detail page and reload the whole page to check persistence.
 
-对已有局的影响（猜测）：6 局都经 `updateTask`、进度组件对空清单返回 null，不会改判。
+Effect on existing runs (estimate): all 6 runs go through `updateTask` and their progress component returns null for an empty checklist, so no verdict changes.
 
 ## related-tasks
 
-没被检查的要求：
-- 「Save links through the API」：判分只在同一浏览器里 `page.reload()`，存进 localStorage 的照样还在。
-- 「lists the other tasks」：选择器里有没有当前任务自己，没查。
+Requirements not checked:
+- "Save links through the API": the judge only does `page.reload()` in the same browser, where links kept in localStorage are still there.
+- "lists the other tasks": whether the picker contains the current task itself was not checked.
 
-| 变体 | 期望 | 判分结果 | 结论 |
+| Variant | Expect | Verdict | Result |
 |---|---|---|---|
-| local-storage（关联只存在 localStorage） | fail | pass | **LEAK** |
-| self-listed（选择器里有当前任务） | fail | pass | **LEAK** |
+| local-storage (links kept only in localStorage) | fail | pass | **LEAK** |
+| self-listed (the picker contains the current task) | fail | pass | **LEAK** |
 | one-way | fail | the link does not show on the other task's page | OK |
-| pass-listbox（自制下拉，role=listbox / option） | pass | pass | OK |
+| pass-listbox (a hand-made dropdown, role=listbox / option) | pass | pass | OK |
 
-建议：刷新前清掉 `localStorage` / `sessionStorage`（或另开一个 context 打开另一边的页面）再核对双向显示；字段名不定，所以不建议去 `api.state()` 里猜字段。选择器：原生 select 读 option 文字，自制的点开后看 option，核对没有「Fix login redirect」（低优先级）。
+Suggested fix: clear `localStorage` / `sessionStorage` before the reload (or open the other task's page in a second context) and then check the link shows both ways; the field name is not fixed, so guessing at fields in `api.state()` is not advised. The picker: for a native select read the option text, for a hand-made one open it and look at the options, and check there is no "Fix login redirect" (low priority).
 
-对已有局的影响（猜测）：6 局都把关联写在任务上经 API 保存、都过滤掉了当前���务，不会改判。
+Effect on existing runs (estimate): all 6 runs write the links onto the task and save through the API, and all filter out the current task, so no verdict changes.
 
 ## shared-task-card
 
-Team Tasks 上的情况：看板卡片本来就有 Star、Mark done/Reopen 和 Delete（✕，原生确认），题面要的「Board 加 Edit 和 Delete」在 Team Tasks 上只剩 Edit 是新的；判分的看板 Delete 检查对什么都不做的看板也成立。
+How it stands on Team Tasks: Board cards already have Star, Mark done / Reopen and Delete (✕, with a native confirm), so of the ticket's "add Edit and Delete to the Board" only Edit is new on Team Tasks; the judge's Board Delete check holds even for a Board that did nothing.
 
-没被检查的要求：
-- 「same dialogs」：`confirmDelete` 有对话框就点、没有就算原生框确认过，看板 Delete 不确认直接删也判通过。
-- 「Nothing else should change」：看板 Star / Unstar、列表行的 Delete、看板拖拽都没查（只查了列表勾选、列表 Edit、看板 Reopen、Mark done、两边没多出对方的控件）。
+Requirements not checked:
+- "same dialogs": `confirmDelete` clicks a dialog if there is one and otherwise counts it as a native dialog having been confirmed, so a Board Delete that deletes without confirming is judged pass too.
+- "Nothing else should change": Star / Unstar on the Board, Delete on a list row and dragging on the Board were never checked (only ticking in the list, Edit in the list, Reopen and Mark done on the Board, and that neither side gained the other's controls).
 
-| 变体 | 期望 | 判分结果 | 结论 |
+| Variant | Expect | Verdict | Result |
 |---|---|---|---|
 | drop-star | fail | pass | **LEAK** |
 | list-delete-gone | fail | pass | **LEAK** |
 | board-delete-no-confirm | fail | pass | **LEAK** |
-| pass-icon-buttons（✎ / 🗑 图标按钮） | pass | pass | OK |
+| pass-icon-buttons (✎ / 🗑 icon buttons) | pass | pass | OK |
 
-建议：
-1. 看板 Delete 和列表 Delete 都要求有确认：点 Delete 后 `ctx.dialogs` 里新增一条 confirm 或者出了应用对话框；最好第一次拒绝、核对任务还在，再确认。
-2. 列表上删一个任务（例如「Rotate API keys」），核对 API 里没了。
-3. 看板卡片上点一次 Star，核对 API 里加了星标，或者至少按钮还在（`/^(un)?star/i`）。
+Suggested fixes:
+1. Require a confirmation for Delete on both the Board and the list: after clicking Delete, `ctx.dialogs` gains a confirm or an app dialog appears; better still, decline the first time, check the task is still there, then confirm.
+2. Delete a task on the list (for example "Rotate API keys") and check it is gone from the API.
+3. Click Star once on a Board card and check the star is set in the API, or at least that the button is still there (`/^(un)?star/i`).
 
-对已有局的影响（猜测）：8 局都保留了 Star、拖拽、列表 Delete 和原生 confirm，不会改判。
+Effect on existing runs (estimate): all 8 runs keep Star, dragging, the list Delete and the native confirm, so no verdict changes.
 
 ## settings-tabs
 
-没被检查的要求：「Split Settings into two tabs」本身：判分只看两个地址能打开、Invites 页列出邀请、Members 页列出成员，两页都照旧显示全部内容的也判通过。
+Requirement not checked: "Split Settings into two tabs" itself. The judge only looks at whether the two addresses open, the Invites page lists invites and the Members page lists members; two pages that both still show everything are judged pass too.
 
-| 变体 | 期望 | 判分结果 | 结论 |
+| Variant | Expect | Verdict | Result |
 |---|---|---|---|
-| no-split（两页都显示成员和邀请） | fail | pass | **LEAK** |
+| no-split (both pages show members and invites) | fail | pass | **LEAK** |
 | invite-button-moved | fail | check threw: locator.click timeout | OK |
-| pass-no-redirect（/settings 直接显示 Members，地址不变） | pass | /settings did not open /settings/members | **STRICT（有争议）** |
-| pass-button-tabs（标签是普通按钮 + navigate） | pass | check threw: locator.click timeout | **STRICT** |
+| pass-no-redirect (/settings shows Members directly, the address does not change) | pass | /settings did not open /settings/members | **STRICT (debatable)** |
+| pass-button-tabs (the tabs are plain buttons + navigate) | pass | check threw: locator.click timeout | **STRICT** |
 
-原因与建议：
-- no-split：在 `/settings/members` 核对邀请邮箱（nina@…）看不见，在 `/settings/invites` 核对「Priya Shah」和「Invite member」看不见。
-- pass-button-tabs：判分找标签只认 `role=tab` 或 main 里的 link；题面只说两个标签各有地址，没说标签用什么元素。建议再接受 main 里名字是 /^invites/i 的 button。
-- pass-no-redirect：题面写「/settings opens the Members tab」，又说两个标签「with their own URLs」，两种读法都说得通（按「题面两种读法都说得通时两种都算对」的原则）：/settings 停在原地址但显示 Members 内容、不显示邀请列表，也判对。
+Causes and suggested fixes:
+- no-split: on `/settings/members` check that the invite's email (nina@…) is not visible, and on `/settings/invites` that "Priya Shah" and "Invite member" are not visible.
+- pass-button-tabs: to find a tab the judge only accepts `role=tab` or a link inside main; the ticket only says the two tabs each have an address, not what element a tab is. Suggested: also accept a button inside main whose name is /^invites/i.
+- pass-no-redirect: the ticket says "/settings opens the Members tab" and also that the two tabs come "with their own URLs"; both readings hold up (by the rule that when a ticket can be read two ways, both count as right): /settings staying at its address while showing the Members content, without the invite list, is judged right too.
 
-对已有局的影响（猜测）：8 局都用 `tab === …` 或嵌套路由分开渲染、都重定向到 /settings/members、都用 NavLink，加严和放宽都不会改判。
+Effect on existing runs (estimate): all 8 runs render the two separately with `tab === …` or nested routes, all redirect to /settings/members and all use NavLink, so neither the stricter nor the looser check changes a verdict.
 
 ## load-retry
 
-没被检查的要求：
-- 接口还在失败时点 Retry：判分先恢复接口再点 Retry，Retry 时把错误清掉、失败后吞掉错误的写法，会再次显示「No tasks here.」，判分看不出来。
-- Retry 是否只是整页刷新：工单抱怨的正是「only way out is reloading the whole page」，判分里刷新整页的按钮也通过（这条有争议）。
+Requirements not checked:
+- Clicking Retry while the API is still failing: the judge restores the API first and then clicks Retry. An implementation that clears the error on Retry and swallows the error when it fails again shows "No tasks here." once more, and the judge cannot see it.
+- Whether Retry is just a full page reload: what the ticket complains about is exactly "only way out is reloading the whole page", yet a button that reloads the page passes the judge (this one is debatable).
 
-| 变体 | 期望 | 判分结果 | 结论 |
+| Variant | Expect | Verdict | Result |
 |---|---|---|---|
-| reload-page（Retry = `location.reload()`） | fail | pass | **LEAK（有争议）** |
+| reload-page (Retry = `location.reload()`) | fail | pass | **LEAK (debatable)** |
 | retry-swallows-error | fail | pass | **LEAK** |
-| pass-reworded（「Something went wrong while loading tasks.」+「Try again」） | pass | no error shown when loading failed | **STRICT** |
+| pass-reworded ("Something went wrong while loading tasks." + "Try again") | pass | no error shown when loading failed | **STRICT** |
 
-建议：
-1. 先在接口仍失败时点一次 Retry，等 1 秒，核对仍有报错、没有「No tasks here.」、Retry 还在；再恢复接口点第二次。
-2. 点 Retry 前 `page.evaluate(() => (window.__noReload = 1))`，Retry 后核对它还在（整页刷新判错）。这条是否算错请定夺。
-3. 「有报错」放宽：报错文字的正则加上 `went wrong|unable|unavailable|problem`，或者 Retry/Try again 按钮出现就算显示了报错。题面没规定报错文字。
+Suggested fixes:
+1. First click Retry once while the API still fails, wait 1 second, and check the error is still shown, there is no "No tasks here." and Retry is still there; then restore the API and click a second time.
+2. Before clicking Retry, `page.evaluate(() => (window.__noReload = 1))`, and after Retry check it is still set (a full reload is wrong). Whether this counts as wrong needs a decision.
+3. Loosen "an error is shown": add `went wrong|unable|unavailable|problem` to the regex for the error text, or count a Retry / Try again button appearing as the error being shown. The ticket does not fix the error text.
 
-对已有局的影响（猜测）：6 局都调 context 的 `reload()`（失败时会重新设 error），不会改判。
+Effect on existing runs (estimate): all 6 runs call the context's `reload()` (which sets error again on failure), so no verdict changes.
 
 ## qa-board-count
 
-判分：报告里有一段同时含 `board|column|to do` 和 `count|number|\b10\b|total|wrong|incorrect|mismatch|doesn't match`。两组词都很宽（`count`、`wrong` 几乎每段都有），而且开头的测试范围、总结段也算一段。
+The judge: some paragraph of the report contains both `board|column|to do` and `count|number|\b10\b|total|wrong|incorrect|mismatch|doesn't match`. Both groups are very broad (`count` and `wrong` are in nearly every paragraph), and the test scope at the top and the summary paragraph count as paragraphs too.
 
-| 变体 | 期望 | 判分结果 | 结论 |
+| Variant | Expect | Verdict | Result |
 |---|---|---|---|
-| no-problems-scope（范围里写「Board (columns, counts …)」，结论 No problems found） | fail | pass | **LEAK** |
-| other-board-bug（只报了看板拖拽时卡片短暂落错列） | fail | pass | **LEAK** |
-| pass-plain-wording（「badge … how many tasks exist overall … both badges are identical」） | pass | report does not say … | **STRICT** |
-| pass-typical（带「says 10 … holds 7 cards」） | pass | pass | OK |
+| no-problems-scope (the scope says "Board (columns, counts …)", the conclusion is No problems found) | fail | pass | **LEAK** |
+| other-board-bug (only reports a card briefly landing in the wrong column while dragging on the Board) | fail | pass | **LEAK** |
+| pass-plain-wording ("badge … how many tasks exist overall … both badges are identical") | pass | report does not say … | **STRICT** |
+| pass-typical (with "says 10 … holds 7 cards") | pass | pass | OK |
 
-建议：只看列表项（以 `-`、`*`、数字开头的段），跳过标题、范围和总结段；报告里有「No problems found」又没有列表项的直接判错。第二组词收窄到这个问题本身的症状：`\b10\b|total|overall|all tasks|identical|same (number|count)|instead of|7\b.*3\b`，去掉单独的 `wrong`、`count`。另外，这两道 QA 题每题只有 6 份报告，人工复核或让模型按题面判比正则可靠。
+Suggested fix: look at list items only (paragraphs starting with `-`, `*` or a number) and skip headings, the scope and the summary; a report that has "No problems found" and no list items is wrong outright. Narrow the second group to the symptoms of this very problem: `\b10\b|total|overall|all tasks|identical|same (number|count)|instead of|7\b.*3\b`, dropping `wrong` and `count` on their own. Besides, each of these two QA tasks has only 6 reports; reviewing them by hand, or having a model judge them against the ticket, is more reliable than regular expressions.
 
-对已有局的影响（猜测）：6 份真实报告都写了列头显示 10、列里实际是 7 张和 3 张，收窄后仍会命中，不会改判。
+Effect on existing runs (estimate): all 6 real reports say the column header shows 10 while the columns really hold 7 and 3 cards; they still match after narrowing, so no verdict changes.
 
 ## qa-priority-ignored
 
-判分：一段里同时含 `priority|high|low` 和 `medium|ignored|always|not (saved|kept|applied|respected)|wrong|instead|default|reset`。`low` 没加词边界，`slow`、`below`、`allow`、`follow` 都会命中；`always`、`default`、`instead`、`wrong` 在别的问题里很常见。
+The judge: one paragraph contains both `priority|high|low` and `medium|ignored|always|not (saved|kept|applied|respected)|wrong|instead|default|reset`. `low` has no word boundary, so `slow`, `below`, `allow` and `follow` all match; `always`, `default`, `instead` and `wrong` are common in other problems.
 
-| 变体 | 期望 | 判分结果 | 结论 |
+| Variant | Expect | Verdict | Result |
 |---|---|---|---|
-| slow-search（只报搜索慢、always 闪 Loading） | fail | pass | **LEAK** |
-| invites-only（只报邀请不能撤销，措辞含 below / allow / instead） | fail | pass | **LEAK** |
-| no-problems-scope（范围里写「sorting … by priority」，结论 No problems found） | fail | pass | **LEAK** |
-| pass-plain-wording（「has no effect … middle level badge」） | pass | report does not say … | **STRICT** |
+| slow-search (only reports that search is slow and always flashes Loading) | fail | pass | **LEAK** |
+| invites-only (only reports that invites cannot be revoked, worded with below / allow / instead) | fail | pass | **LEAK** |
+| no-problems-scope (the scope says "sorting … by priority", the conclusion is No problems found) | fail | pass | **LEAK** |
+| pass-plain-wording ("has no effect … middle level badge") | pass | report does not say … | **STRICT** |
 
-真实文字的旁证：拿别的 QA 题的真实报告来套这道题的判分，Mirasim Computer Use 第 2、3 局的 qa-header-shadow 报告（报的是页头遮挡点击、邀请不能撤销、成员数不刷新，和优先级无关）都会判「发现了优先级被忽略」：命中的是「allowed」（含 low）+「At the default 1440×813 viewport」、「below the header」+「default desktop viewport」。也就是说，现实里模型漏掉这个 bug、报了别的问题，也大概率判通过。
+Evidence from real text: run through this task's judge, the real reports of another QA task also pass. The qa-header-shadow reports of Mirasim Computer Use runs 2 and 3 (about the header covering clicks, invites that cannot be revoked and a member count that does not refresh, nothing to do with priority) are both judged "found the priority being ignored": the matches are "allowed" (contains low) + "At the default 1440×813 viewport", and "below the header" + "default desktop viewport". In other words, a model that misses this bug in practice and reports other problems most likely passes as well.
 
-建议：第一组改成 `\bpriority\b|\bhigh\b|\blow\b`，并要求同一段里提到新建（`add|new|creat`）；第二组去掉 `always|wrong|instead`，保留 `medium|ignored|not (saved|kept|applied|respected)|default|reset`，再加上 `no effect|middle`。同样��看列表项、跳过范围和总结段。
+Suggested fix: the first group becomes `\bpriority\b|\bhigh\b|\blow\b`, and the same paragraph must mention creating (`add|new|creat`); drop `always|wrong|instead` from the second group, keep `medium|ignored|not (saved|kept|applied|respected)|default|reset` and add `no effect|middle`. As above, look at list items only and skip the scope and the summary.
 
-对已有局的影响（猜测）：6 份真实报告说的都是新建任务时选的优先级被忽略、显示成 Medium，收窄后仍命中，不会改判。
+Effect on existing runs (estimate): all 6 real reports say the priority chosen when creating a task is ignored and shows as Medium; they still match after narrowing, so no verdict changes.
 
 ## undo-delete
 
-没被检查的要求：
-- 「for a few seconds」：没查提示会消失。
-- 「Undo brings the task back」：只按标题核对，Undo 用 POST 只按标题重建的写法（优先级、完成状态、描述都丢了）也通过。
-- 判分在不点 Undo 的删除后固定等 9 秒就读 API，撤销窗口超过约 8 秒的实现会被判「never reached the API」。
+Requirements not checked:
+- "for a few seconds": the notice going away was not checked.
+- "Undo brings the task back": only the title is compared, so an Undo that POSTs a new task from the title alone (priority, done state and description all lost) passes too.
+- After a delete without Undo the judge waits a fixed 9 seconds and then reads the API, so an implementation whose undo window is longer than about 8 seconds is judged "never reached the API".
 
-| 变体 | 期望 | 判分结果 | 结论 |
+| Variant | Expect | Verdict | Result |
 |---|---|---|---|
-| toast-sticks（提示一直不消失） | fail | pass | **LEAK** |
-| undo-title-only（立即删，Undo 只按标题重建） | fail | pass | **LEAK** |
-| pass-10s-window（撤销窗口 10 秒） | pass | a delete without Undo never reached the API | **STRICT** |
-| pass-recreate-full（立即删，Undo 带全部字段重建） | pass | pass | OK |
+| toast-sticks (the notice never goes away) | fail | pass | **LEAK** |
+| undo-title-only (deletes at once, Undo recreates from the title only) | fail | pass | **LEAK** |
+| pass-10s-window (a 10-second undo window) | pass | a delete without Undo never reached the API | **STRICT** |
+| pass-recreate-full (deletes at once, Undo recreates with every field) | pass | pass | OK |
 
-建议：
-1. 最后那次不点 Undo 的删除，改成轮询：最多等约 20 秒，等到 API 里删掉（或打了删除标记）就继续；同时核对「Task deleted」在这段时间内消失（toast-sticks 判错）。这样 10 秒的窗口也能过。
-2. Undo 之后核对 API 里「Update team wiki」仍是 `done: true`、优先级仍是 low，「Plan team offsite」的优先级和描述没变（undo-title-only 判错；带全部字段重建的仍判对）。
+Suggested fixes:
+1. For the last delete, the one without Undo, poll instead: wait up to about 20 seconds and go on as soon as the task is deleted in the API (or marked deleted); also check that "Task deleted" goes away within that time (toast-sticks is wrong). A 10-second window then passes too.
+2. After Undo, check in the API that "Update team wiki" is still `done: true` with priority low, and that the priority and description of "Plan team offsite" are unchanged (undo-title-only is wrong; recreating with every field is still right).
 
-对已有局的影响（猜测）：8 局都是 5 秒窗口、到时才真删、原记录不动，提示 5 秒后消失，不会改判。
+Effect on existing runs (estimate): all 8 runs have a 5-second window, really delete only when it ends, leave the original record alone and take the notice away after 5 seconds, so no verdict changes.
 
 ---
 
-## 最该先改的 5 处
+## The five things to fix first
 
-1. **两道 QA 题的报告匹配**（qa-priority-ignored 最急，qa-board-count 其次）：「No problems found」加一行带关键词的测试范围就能判通过；qa-priority-ignored 的 `low` 能匹配 slow、below、allow，Mirasim Computer Use 第 2、3 局真实的 qa-header-shadow 报告套上去都会误判。只看列表项，收窄关键词，或者干脆人工、模型复核（每题 6 份）。
-2. **undo-delete**：提示不消失、Undo 只按标题重建都判通过；固定等 9 秒又会冤枉撤销窗口稍长的实现。改成轮询等删除、核对提示消失、核对恢复后的字段。
-3. **shared-task-card 的「其他不变」**：看板丢了 Star、列表丢了 Delete、看板 Delete 不确认都判通过；而 Team Tasks 看板本来就有 Delete，这道题现在实际只考 Edit。
-4. **invite-admin**：撤销前要先确认、没有邀请时只显示「Settings」都没查；计数从未在页面重新打开后看过。
-5. **related-tasks 的「经 API 保存」**：同一浏览器 reload 分不出 localStorage；刷新前清存储或换 context。同类问题还有 checklist（刷新后才看列表进度，`stale-list`），改法相同：先在客户端跳转后��对，再整页刷新或直接打开别的页核对。
+1. **Report matching in the two QA tasks** (qa-priority-ignored most urgent, qa-board-count next): "No problems found" plus one line of test scope with keywords in it is judged pass; the `low` of qa-priority-ignored matches slow, below and allow, and the real qa-header-shadow reports of Mirasim Computer Use runs 2 and 3 are judged wrongly when run through it. Look at list items only and narrow the keywords, or simply review by hand or by model (6 reports per task).
+2. **undo-delete**: a notice that never goes away and an Undo that recreates from the title alone are both judged pass; the fixed 9-second wait in turn wrongly fails implementations with a slightly longer undo window. Poll for the delete, check the notice goes away, check the fields after the restore.
+3. **"Nothing else changes" in shared-task-card**: a Board that lost Star, a list that lost Delete and a Board Delete without a confirm are all judged pass; and since the Team Tasks Board already has Delete, this task now really only tests Edit.
+4. **invite-admin**: confirming before a revoke and showing just "Settings" when there are no invites were not checked; the count was never looked at after the page was opened again.
+5. **"Saved through the API" in related-tasks**: a reload in the same browser cannot tell localStorage apart; clear the storage before the reload or use another context. checklist has the same kind of problem (list progress only looked at after a reload, `stale-list`) and the same fix: check after client-side navigation first, then reload the whole page or open another page directly.
 
-STRICT 里值得一起改的：load-retry 的报错文字正则（`pass-reworded`）、settings-tabs 标签只认 tab/link（`pass-button-tabs`）、undo-delete 的 9 秒（见第 2 条）。invite-admin 的「Settings 2」徽标、settings-tabs 的「/settings 不重定向」两处请定夺。
+STRICT findings worth fixing along the way: the error-text regex of load-retry (`pass-reworded`), settings-tabs accepting only tab/link for a tab (`pass-button-tabs`), the 9 seconds of undo-delete (see item 2). The "Settings 2" badge of invite-admin and "/settings does not redirect" of settings-tabs need a decision.
 
 ---
 
-## 修订结果
+## What was changed
 
-按上面的建议全部改了。沿用 classic 题（上）的做法：改动全在 `tasks/classic-on-team-tasks.ts`（每题换一个只在 Team Tasks 上用的检查），`tasks/classic.ts` 没动，classic 题在 Team Tasks 前一版上的判分不变。四处有争议的按定夺：invite-admin「Settings 2」算对、settings-tabs /settings 不重定向算对、load-retry 的 Retry 整页刷新算错（why.md 已改成「已定」）。新加的共用辅助函数：`askThenConfirm`（先点一次并拒绝，要求没删；再点一次确认），`navCount`（导航计数）。
+Everything was changed as suggested above. As in part 1, the changes are all in `tasks/classic-on-team-tasks.ts` (each task gets a check used only on Team Tasks); `tasks/classic.ts` was not touched, so the judging of classic tasks on the previous version of Team Tasks is unchanged. The four debatable points went as decided: "Settings 2" in invite-admin is right, /settings without a redirect in settings-tabs is right, and a Retry that reloads the whole page in load-retry is wrong (their why.md now says "decided"). New shared helpers: `askThenConfirm` (click once and decline, requiring that nothing was deleted; click again and confirm) and `navCount` (the count in the nav).
 
-每题的步骤：`harness/selftest.ts -j 1 <id>`（种子判失败、参考答案判通过）→ `harness/audit.ts -j 1 <id>` → 用只读的重判脚本（不写回结果文件）把这题 v1 正式跑里两侧的局全部重判一遍。9 题全部改了。
+Steps for each task: `harness/selftest.ts -j 1 <id>` (the seed is judged fail, the reference solution pass) → `harness/audit.ts -j 1 <id>` → re-judge all of the task's v1 runs on both sides with the read-only re-judge script (which does not write back to the result files). All 9 tasks were changed.
 
-| 题 | 收紧 | 放宽 | 自检 | 变体 | 只读重判 |
+| Task | Stricter | Looser | Self-test | Variants | Read-only re-judge |
 |---|---|---|---|---|---|
-| invite-admin | 撤销前先确认（先拒绝一次邀请要还在）；0 个邀请时只写 Settings；整页重新打开后计数还在 | 计数不必带括号 | OK | 6/6 | 两侧第 1–3 局全过 |
-| checklist | 改完清单不刷新就用侧栏去列表、看板看 1/2；看板上没清单的任务不显示进度 | 进度认「1 / 2」「1 of 2」 | OK | 4/4 | 两侧第 1–3 局全过 |
-| related-tasks | 清掉 localStorage / sessionStorage 再刷新看另一边；选择器里不能有当前任务 | — | OK | 4/4 | 两侧第 1–3 局全过 |
-| shared-task-card | 列表和看板的 Delete 都先确认；列表 Delete 还删得掉；看板 Star 还在、点了存进 API | — | OK（参考答案见下） | 4/4 | 两侧第 1–3 局和 Mirasim Computer Use 两局试跑全过 |
-| settings-tabs | Members 页看不到待处理邀请，Invites 页看不到成员 | /settings 停在原地址显示 Members 也算；标签认 main 里的 Invites 按钮 | OK | 4/4 | 两侧第 1–3 局和各一局试跑全过 |
-| load-retry | 接口仍失败时点 Retry 要照样报错、不说没有任务；Retry 不能整页刷新 | 报错不限措辞（went wrong / unable / unavailable / problem，或出了 Retry 按钮）；「没有任务」按整词 no tasks | OK | 3/3 | 两侧第 1–3 局全过 |
-| qa-board-count | 按条目、整词判；「No problems found」和范围说明不算，否定句不算；条目要说到看板和列头，某句要说出数字不对 | 不必出现 count / 10 等词（「badge … overall … identical」也算） | OK | 4/4 | 见下 |
-| qa-priority-ignored | 同上；条目要整词说到优先级和新建，某句要说出没生效（显示 / 存成 medium，「at medium priority」这种描述输入的不算） | 「has no effect」「middle level」也算 | OK | 4/4 | 见下 |
-| undo-delete | Undo 后 priority、done、description 和删除前一样；不点 Undo 时提示 20 秒内自己消失 | 不点 Undo 的删除不再固定等 9 秒，最多等 20 秒、删到后端就继续 | OK | 4/4 | 两侧第 1–3 局和 Mirasim Computer Use 两局试跑全过 |
-
-参考答案改动：shared-task-card 原来的参考答案里 BoardPage 是照 Team Tasks 的前一版写的，盖到 Team Tasks 上会丢掉看板的 Star 和拖拽，按新判分不符合题面。新加了 Team Tasks 上的参考答案 `references/shared-task-card/`（取自 v1 正式跑 ChatGPT Computer Use 第 2 局的产出：共用 `TaskItem` 组件，看板保留 Star、拖拽、Mark done / Reopen，加 Edit 对话框），之后自检通过。
-
-QA 两题的交叉测：拿历次跑的局（更早几批、Team Tasks 前一版上的局、v1 正式跑）和变体里的全部 151 份 qa-report.md 套两个新判分。
-- 别题的报告（qa-overlay、qa-header-shadow、qa-members-error，以及两题互相）没有一份误中。第一版 qa-priority-ignored 误中了 2 份（更早一批 qa-overlay 题的一份报告写了「reproduced … at medium priority」，Team Tasks 前一版上 Mirasim Computer Use 第 1 局的 qa-header-shadow 报告写了「at the normal 1440×813 viewport」），已改成 medium / normal 前面要有「显示 / 存成」这类动词才算，再测 0 误中。
-- 本题在 Team Tasks 前一版上的真实报告（两侧各 8 份）和 v1 正式跑的真实报告（两侧各 6 份）全部命中，所以 v1 正式跑的 12 局按新判分仍全部通过（这两题的判分只读报告，等同重判）。
-- 合成的否定句（「counts are correct」「No issue with the Board column counts」「don't look wrong」「the priority is not ignored」「works as expected」）和「范围说明 + No problems found」都判未通过。
-
-最后把 9 题一起重跑：自检 9/9 OK，变体 37/37 OK（LEAK 0、STRICT 0）。
-
-**结论变化：** 只读重判的 v1 正式跑局（9 题共 60 局：7 道功能 / 修复题 48 局，加两道 QA 题的 12 份报告）没有一局结论变化。
+| invite-admin | Confirm before a revoke (after declining once the invite must still be there); only Settings at 0 invites; the count is still there after the whole page is opened again | The count need not be in brackets | OK | 6/6 | Runs 1–3 on both sides all pass |
+| checklist | After changing the checklist, go to the list and the Board through the sidebar without a reload and look for 1/2; no progress on the Board for tasks without a checklist | "1 / 2" and "1 of 2" are accepted as progress | OK | 4/4 | Runs 1–3 on both sides all pass |
+| related-tasks | Clear localStorage / sessionStorage, then reload and look at the other side; the picker must not contain the current task | — | OK | 4/4 | Runs 1–3 on both sides all pass |
+| shared-task-card | Delete confirms first on both the list and the Board; Delete on the list still deletes; Star on the Board is still there and a click is saved to the API | — | OK (for the reference solution see below) | 4/4 | Runs 1–3 on both sides and the two Mirasim Computer Use trial runs all pass |
+| settings-tabs | Pending invites are not visible on the Members page, members are not visible on the Invites page | /settings staying at its address and showing Members counts too; an Invites button inside main is accepted as a tab | OK | 4/4 | Runs 1–3 on both sides and one trial run each all pass |
+| load-retry | Retry while the API still fails must show the error again and not say there are no tasks; Retry must not reload the whole page | Any wording for the error (went wrong / unable / unavailable / problem, or a Retry button appearing); "no tasks" as the whole words no tasks | OK | 3/3 | Runs 1–3 on both sides all pass |
+| qa-board-count | Judged per entry and per whole word; "No problems found" and a statement of scope do not count, negated sentences do not count; the entry must mention the Board and the column header, and some sentence must say the number is wrong | Words such as count / 10 need not appear ("badge … overall … identical" counts too) | OK | 4/4 | See below |
+| qa-priority-ignored | Same; the entry must mention priority and creating in whole words, and some sentence must say it had no effect (shown or saved as medium; a description of the input such as "at medium priority" does not count) | "has no effect" and "middle level" count too | OK | 4/4 | See below |
+| undo-delete | After Undo, priority, done and description are as before the delete; without Undo the notice goes away by itself within 20 seconds | A delete without Undo is no longer given a fixed 9 seconds: the judge waits up to 20 seconds and goes on once the delete reaches the backend | OK | 4/4 | Runs 1–3 on both sides and the two Mirasim Computer Use trial runs all pass |
+
+Reference solution change: in the old reference solution of shared-task-card, BoardPage was written against the previous version of Team Tasks; laid over Team Tasks it drops the Board's Star and dragging, which does not meet the ticket under the new judge. A reference solution for Team Tasks was added, `references/shared-task-card/` (taken from the output of ChatGPT Computer Use run 2 of the v1 runs: a shared `TaskItem` component, the Board keeping Star, dragging and Mark done / Reopen, plus an Edit dialog), and the self-test then passes.
+
+Cross-test of the two QA tasks: all 151 qa-report.md files from runs over time (earlier batches, runs on the previous version of Team Tasks, the v1 runs) and from the variants were run through the two new judges.
+- No report of another task (qa-overlay, qa-header-shadow, qa-members-error, or either task against the other) matches wrongly. The first version of qa-priority-ignored matched 2 wrongly (a report of the qa-overlay task from an earlier batch says "reproduced … at medium priority", and the qa-header-shadow report of Mirasim Computer Use run 1 on the previous version of Team Tasks says "at the normal 1440×813 viewport"); it now requires a verb such as "shows / saved as" before medium / normal, and a retest has 0 wrong matches.
+- This task's real reports on the previous version of Team Tasks (8 per side) and those of the v1 runs (6 per side) all match, so the 12 runs of the v1 runs all still pass under the new judges (these two judges only read the report, so this amounts to a re-judge).
+- Made-up negated sentences ("counts are correct", "No issue with the Board column counts", "don't look wrong", "the priority is not ignored", "works as expected") and "a statement of scope + No problems found" are all judged not passed.
+
+Finally all 9 tasks were run again together: self-test 9/9 OK, variants 37/37 OK (LEAK 0, STRICT 0).
+
+**Verdict changes:** of the v1 runs re-judged read-only (60 runs over the 9 tasks: 48 runs of the 7 feature and fix tasks, plus the 12 reports of the two QA tasks), none changed verdict.

docs/judge-audits/controls.md

@@ -1,223 +1,223 @@
-# 判分审计：控件题（9 道 extended 题）
+# Judge audit: control tasks (9 extended tasks)
 
-本文只保留 v1 题单里的题。
+Only tasks in the v1 suite are kept here.
 
-题：board-dnd、assignee-combo、due-off-by-one、infinite-dup、row-actions-kbd、draft-keep、attach-upload、back-keeps-place、old-links。
+Tasks: board-dnd, assignee-combo, due-off-by-one, infinite-dup, row-actions-kbd, draft-keep, attach-upload, back-keeps-place, old-links.
 
-变体都在 `audit-variants/<题id>/<变体名>/`，`npx tsx harness/audit.ts -j 1 <题id>` 复跑。修 bug 题的变体在打了种子补丁的 Team Tasks 上改；assignee-combo、draft-keep 以 `references/<id>/` 为底（`files/`）；attach-upload、old-links 以不打种子的 Team Tasks 原文件为底（`files/`），`edits.json` 只改一处。共 34 个变体：**LEAK 15，STRICT 5（4 个不同的原因）**，其余 14 个 OK，没有 ERR。
+The variants are in `audit-variants/<id>/<variant>/`; rerun them with `npx tsx harness/audit.ts -j 1 <id>`. Variants of bug-fix tasks are edits to Team Tasks with the seed patch applied; assignee-combo and draft-keep build on `references/<id>/` (`files/`); attach-upload and old-links build on the original Team Tasks files without the seed (`files/`), with `edits.json` changing one place. 34 variants in all: **LEAK 15, STRICT 5 (4 different causes)**, the other 14 OK, no ERR.
 
-除了跑变体，还把每个 LEAK 对应的「建议加进判分的检查」写成探针脚本（不在仓库里），在三类代码上各跑了一遍：参考答案（全部通过）、变体（LEAK 的变体全部被抓住，expect pass 的变体全部通过）、**9 题全部 64 局真实产出**（100 次探针全部通过）。所以下面的 LEAK 修法加进判分后，按探针的结果，已有局都不会改判；这一条是实测的，但正式改判分时写法会和探针不同，仍需按计划重判确认。
+Besides running the variants, the "check to add to the judge" suggested for each LEAK was written as a probe script (not in the repository) and run on three kinds of code: the reference solutions (all pass), the variants (every LEAK variant is caught, every expect-pass variant passes) and **all 64 real runs of the 9 tasks** (all 100 probe runs pass). So by the probes, adding the LEAK fixes below to the judges changes no verdict of an existing run. This was measured, but the real judge changes will be written differently from the probes and still need the planned re-judge to confirm it.
 
-## 总表
+## Summary
 
-| 题 id | 要求条数 | 没被检查的要求 | 变体数 | LEAK | STRICT | 真实产出里的误判 |
+| Task id | Requirements | Requirements not checked | Variants | LEAK | STRICT | Wrong verdicts on real runs |
 |---|---|---|---|---|---|---|
-| board-dnd | 7 | Done → To do 保存失败时的回滚；拖回本列 / 列外松手不该移动 | 4 | 2 | 0 | 无 |
-| assignee-combo | 8 | 方向键（判分每次输入都只剩一个匹配，方向键没被用到）；输入名字中间的部分；读屏能知道高亮的是哪一项 | 4 | 3 | 0 | 无 |
-| due-off-by-one | 5 | 西边时区（西雅图）的编辑页回显和「再存一次不挪」（只在上海查） | 3 | 1 | 0 | 无 |
-| infinite-dup | 3（+1 隐含） | 与后端快慢无关（只在每页 1 秒下查）；隐含：加载中换筛选后仍能滚到底 | 3 | 2 | 0 | 无 |
-| row-actions-kbd | 5 | 其他行保持隐藏（只看焦点所在那行） | 4 | 1 | 1 | 无 |
-| draft-keep | 5（+1 隐含） | 「在这个浏览器里」（换个标签页还在）；隐含：发送失败不丢草稿 | 4 | 2 | 0 | 无 |
-| attach-upload | 6 | 再点一次文件名收起预览；拒绝过大文件后还能照常上传 | 5 | 2 | 2（同一原因） | **Mirasim Computer Use 第 1 局误判未通过** |
-| back-keeps-place | 5 | 「完全」停在原处（48px 容差 + 挑视口中段的行，把那行居中也能过） | 3 | 1 | 1 | 无 |
-| old-links | 9 | 已删除提示里「谁删的、何时删的」 | 4 | 1 | 1（有争议） | 无 |
-| **合计** | | | **34** | **15** | **5** | **1** |
+| board-dnd | 7 | Rollback when a Done → To do save fails; a drop back into the same column or outside the columns should not move the card | 4 | 2 | 0 | none |
+| assignee-combo | 8 | Arrow keys (every input the judge types leaves one match, so the arrow keys are never used); typing a middle part of a name; a screen reader knowing which option is highlighted | 4 | 3 | 0 | none |
+| due-off-by-one | 5 | In a western time zone (Seattle), what the edit page shows and "saving again does not shift it" (only checked in Shanghai) | 3 | 1 | 0 | none |
+| infinite-dup | 3 (+1 implied) | Independence from backend speed (only checked at 1 second per page); implied: the list still scrolls to the end after the filter changes mid-load | 3 | 2 | 0 | none |
+| row-actions-kbd | 5 | Other rows staying hidden (only the focused row is looked at) | 4 | 1 | 1 | none |
+| draft-keep | 5 (+1 implied) | "In this browser" (still there in another tab); implied: a failed send does not lose the draft | 4 | 2 | 0 | none |
+| attach-upload | 6 | Clicking the file name again hides the preview; uploads still work after an oversized file is refused | 5 | 2 | 2 (same cause) | **Mirasim Computer Use run 1 wrongly judged not passed** |
+| back-keeps-place | 5 | Staying "exactly" where it was (a 48px tolerance plus picking a row in the middle of the viewport lets centring that row pass) | 3 | 1 | 1 | none |
+| old-links | 9 | "Who deleted it and when" in the deleted-task notice | 4 | 1 | 1 (debatable) | none |
+| **Total** | | | **34** | **15** | **5** | **1** |
 
-「要求条数」按工单逐句拆（多轮题各轮相加），「隐含」是题面没写、但 Team Tasks 原本就有，改坏了算回归的行为。
+"Requirements" counts the ticket sentence by sentence (the rounds are added up for multi-round tasks). "Implied" is behaviour the ticket does not state but Team Tasks already had, so breaking it is a regression.
 
-真实产出：9 题两边共 64 局（每题每边第 1–3 局；board-dnd、assignee-combo 两边各多一局试跑；old-links 的 ChatGPT Computer Use 一侧另有两局试跑，Mirasim Computer Use 一侧另有四局试跑），除 attach-upload 的 Mirasim Computer Use 第 1 局外全部判通过。读了每题 2–6 局的 diff（board-dnd 4 局、assignee-combo 8 局的新组件、due-off-by-one 2 局、infinite-dup 6 局、row-actions-kbd 2 局、draft-keep 6 局、attach-upload 6 局、back-keeps-place 2 局、old-links 4 局），再用上面的探针把 64 局全跑了一遍：**没有「判了通过但其实没做到题面」的局**；有 **1 局判了未通过其实是对的**（attach-upload 的 Mirasim Computer Use 第 1 局，见下）。这批产出彼此很像：board-dnd 都是 `elementsFromPoint` 跳过卡片或 `pointer-events: none` 加 `finally`；row-actions-kbd 6 局都是去掉 `tabIndex={-1}` 加 `:focus-within`（带 `pointer-events: auto`）；due-off-by-one 6 局都是改回 `f.dueDate || null`。下面这些漏洞眼下没造成误判通过，但判分对这些错法没有防护。
+Real runs: 64 runs over both sides of the 9 tasks (runs 1–3 per side per task; one more trial run per side for board-dnd and assignee-combo; for old-links two more trial runs on the ChatGPT Computer Use side and four more on the Mirasim Computer Use side), all judged pass except Mirasim Computer Use run 1 of attach-upload. The diffs of 2–6 runs per task were read (board-dnd 4 runs, the new component of 8 assignee-combo runs, due-off-by-one 2, infinite-dup 6, row-actions-kbd 2, draft-keep 6, attach-upload 6, back-keeps-place 2, old-links 4), and then the probes above were run over all 64 runs: **no run was judged pass without doing what the ticket asks**; **1 run was judged not passed although it is right** (Mirasim Computer Use run 1 of attach-upload, see below). These runs are much alike: board-dnd is always `elementsFromPoint` skipping the card, or `pointer-events: none` with a `finally`; all 6 row-actions-kbd runs remove `tabIndex={-1}` and add `:focus-within` (with `pointer-events: auto`); all 6 due-off-by-one runs go back to `f.dueDate || null`. The holes below have caused no wrongly passed run so far, but the judges have no defence against these mistakes.
 
 ---
 
 ## board-dnd
 
-要求：①To do → Done 拖了能保存；②Done → To do 也能；③卡片显示在新列；④每张卡都行；⑤保存失败有提示；⑥失败后卡片回到原来那列；⑦（拖到另一列才算移动——工单「carry it over to the other column」）。判分查了 ①②③⑤，④ 查了 3 张卡，⑥ 只查了 To do → Done 方向，⑦ 没查。另外查了失败后还能再拖、点卡片还能打开。
+Requirements: ① a drag from To do → Done is saved; ② so is Done → To do; ③ the card shows in the new column; ④ it works for every card; ⑤ a failed save shows a notice; ⑥ after a failure the card goes back to the column it came from; ⑦ (only a drop on the other column counts as a move: the ticket says "carry it over to the other column"). The judge checked ①②③⑤, ④ on 3 cards, ⑥ only in the To do → Done direction, and not ⑦. It also checked that dragging still works after a failure and that clicking a card still opens it.
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
 | drop-toggles | fail | **LEAK** |
 | revert-hardcoded-todo | fail | **LEAK** |
-| only-round1 | fail | OK（after a failed save the card stays in Done…） |
+| only-round1 | fail | OK (after a failed save the card stays in Done…) |
 | no-optimistic | pass | OK |
 
-- **drop-toggles**（LEAK）：根本没修落点判断（拖动中的卡仍挡住 `elementFromPoint`，落点永远是原列），改成「只要拖过就翻到另一列」。判分的三次拖都是跨列，全过；但把卡拖回本列或拖到列外松手，卡片照样被挪走并保存。探针「把 Write release notes 拖回 To do，1.5 秒后仍未完成」在参考答案和 8 局真实产出上都过，在本变体上失败。
-  建议：判分开头加一步「拖回本列松手 → 任务 1 仍是未完成、卡片仍在 To do」。
-- **revert-hardcoded-todo**（LEAK）：失败时把 pending 写死成 `done: false`。To do → Done 失败看着对；Done → To do 失败时卡片留在 To do（放下的那列），没回 Done。判分的失败场景只有任务 8（在 To do）。
-  建议：判分场景再让一张 Done 卡（如任务 6）的 PATCH 也失败，拖到 To do 后应仍在 Done；现有任务 6 那一步要换一张 Done 卡做成功的拖动。探针（任务 6 失败、拖到 To do、2.5 秒后仍在 Done）在参考答案和 8 局上都过。
-- 影响已有局：探针 16 次全过，猜测不会改判。
+- **drop-toggles** (LEAK): does not fix where the drop lands at all (the dragged card still blocks `elementFromPoint`, so the drop target is always the original column) and instead "flips the card to the other column whenever it was dragged". The judge's three drags all cross columns and all pass; but dragging a card back into its own column, or releasing it outside the columns, still moves and saves it. The probe "drag Write release notes back into To do; 1.5 seconds later it is still not done" passes on the reference solution and on the 8 real runs and fails on this variant.
+  Suggested fix: add a step at the start of the judge, "drop back into the same column → task 1 is still not done and the card is still in To do".
+- **revert-hardcoded-todo** (LEAK): on failure pending is hard-coded to `done: false`. A failed To do → Done looks right; after a failed Done → To do the card stays in To do (the column it was dropped on) and does not return to Done. The judge's failure scenario only has task 8 (in To do).
+  Suggested fix: have the judging scenario also fail the PATCH of a Done card (such as task 6); dragged to To do it should still be in Done. The existing step with task 6 then needs another Done card for the successful drag. The probe (task 6 fails, dragged to To do, still in Done after 2.5 seconds) passes on the reference solution and on the 8 runs.
+- Effect on existing runs: all 16 probe runs pass; the estimate is no verdict changes.
 
 ## assignee-combo
 
-要求：①输入名字的一部分，列表缩到匹配项；②方向键在匹配项间移动；③回车选高亮项且不提交整个表单；④Esc 关列表、不改选择；⑤鼠标点匹配项也行；⑥Unassigned 仍可选；⑦标签仍是 Assignee；⑧像标准 combobox 一样给键盘和读屏用户用。判分查了 ③④⑤⑥⑦；① 只输入名字开头（dieg、pri、mei、unass）；② 实际没查——这四个输入都只剩一个匹配，判分循环里高亮一开始就在目标上，一次方向键都不按；⑧ 只查了 role=combobox / option，高亮可以只靠类名。
+Requirements: ① typing part of a name narrows the list to the matches; ② the arrow keys move between matches; ③ Enter picks the highlighted one and does not submit the whole form; ④ Esc closes the list without changing the choice; ⑤ clicking a match with the mouse works too; ⑥ Unassigned can still be chosen; ⑦ the label is still Assignee; ⑧ usable by keyboard and screen-reader users like a standard combobox. The judge checked ③④⑤⑥⑦; ① only with the start of a name (dieg, pri, mei, unass); ② in fact not at all: each of those four inputs leaves one match, the highlight in the judge's loop starts on the target, and no arrow key is ever pressed; ⑧ only role=combobox / option, and the highlight may rest on a class name alone.
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
 | arrows-ignored | fail | **LEAK** |
 | prefix-only | fail | **LEAK** |
 | sr-blind | fail | **LEAK** |
 | no-auto-highlight | pass | OK |
 
-- **arrows-ignored**（LEAK）：方向键什么都不做，回车总选第一个匹配。
-  建议：加一步输入能匹配多人的字（如 "a"），按 ↓↓↑↓ 后回车，选中的不应是第一个匹配项（探针就是这样写的；不管初始高亮在第一项还是没有高亮，正确实现都会落在第 2 或第 3 项）。
-- **prefix-only**（LEAK）：过滤用 `startsWith(全名)`，输入姓 "alvar" 找不到 Diego Alvarez。题面是「type part of a person's name」。
-  建议：把其中一次输入换成姓的一部分（如 "alvar" 或 "shah"）。
-- **sr-blind**（LEAK）：去掉 `aria-activedescendant`、所有 option 的 `aria-selected` 都是 false，高亮只靠 `.active` 类。读屏听不到当前项。判分 `highlightedOption` 最后一档按类名认。
-  建议：按过方向键后，要求焦点在 option 上，或输入框的 `aria-activedescendant` 指向一个 option（读屏能感知的两种标准做法），去掉「唯一一个 active 类」那一档（`aria-selected` 那一档可留）。8 局真实产出都用了 `aria-activedescendant`。
-- no-auto-highlight（打字后不自动高亮，按 ↓ 才高亮）判通过，判分对这种合法写法不严；两边各一局试跑就是这种写法。
-- 影响已有局：三个探���在 8 局上 24 次全过，猜测不会改判。
+- **arrows-ignored** (LEAK): the arrow keys do nothing and Enter always picks the first match.
+  Suggested fix: add a step that types something matching several people (such as "a"), presses ↓↓↑↓ and then Enter; the pick should not be the first match (the probe is written this way; whether the highlight starts on the first option or nowhere, a correct implementation ends on the 2nd or 3rd).
+- **prefix-only** (LEAK): filters with `startsWith(full name)`, so typing the surname "alvar" does not find Diego Alvarez. The ticket says "type part of a person's name".
+  Suggested fix: change one of the inputs to part of a surname (such as "alvar" or "shah").
+- **sr-blind** (LEAK): `aria-activedescendant` is removed and `aria-selected` is false on every option; the highlight rests on the `.active` class alone. A screen reader does not announce the current option. The last tier of the judge's `highlightedOption` accepts a class name.
+  Suggested fix: after arrow keys were pressed, require focus on an option or the input's `aria-activedescendant` pointing at an option (the two standard ways a screen reader can tell), and drop the "the only active class" tier (the `aria-selected` tier can stay). All 8 real runs use `aria-activedescendant`.
+- no-auto-highlight (no highlight after typing, only after ↓) is judged pass; the judge is not strict on this legitimate approach, which is what one trial run on each side does.
+- Effect on existing runs: the three probes pass all 24 times on the 8 runs; the estimate is no verdict changes.
 
 ## due-off-by-one
 
-要求：①存下的是选的那天；②任务上显示那天；③编辑页显示那天；④改别的字段再存不挪；⑤上海、柏林、西雅图的人都对。判分在上海查了 ①②③④，在西雅图只查了 ①②。柏林没查，但与上海同在 UTC 以东，出错方式相同，不算缺口。
+Requirements: ① the day picked is the day saved; ② the task shows that day; ③ the edit page shows that day; ④ changing another field and saving again does not shift it; ⑤ right for people in Shanghai, Berlin and Seattle. The judge checked ①②③④ in Shanghai and only ①② in Seattle. Berlin is not checked, but it lies east of UTC like Shanghai and goes wrong the same way, so that is no gap.
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
 | form-utc-west | fail | **LEAK** |
-| plus-one-day | fail | OK（in Seattle, picking Dec 1 saved "2026-12-02"） |
+| plus-one-day | fail | OK (in Seattle, picking Dec 1 saved "2026-12-02") |
 | display-utc | pass | OK |
 
-- **form-utc-west**（LEAK）：保存修对了，但编辑页读回日期用 `new Date("YYYY-MM-DD").toLocaleDateString("en-CA")`。上海正确；西雅图编辑页显示 11-30，再存一次就挪到前一天——正是工单里 Mei「改两次挪两天」的现象。
-  建议：西雅图那段也重开编辑页查回显，再改 Notes 存一次查日期不变（和上海那段对称）。
-- 影响已有局：6 局都只改了保存那一行，探针 6 次全过，猜测不会改判。
+- **form-utc-west** (LEAK): saving is fixed, but the edit page reads the date back with `new Date("YYYY-MM-DD").toLocaleDateString("en-CA")`. Right in Shanghai; in Seattle the edit page shows 11-30 and another save shifts the date a day earlier, which is exactly Mei's "edit twice, two days off" in the ticket.
+  Suggested fix: in the Seattle part too, reopen the edit page and check what it shows, then change Notes, save and check the date is unchanged (mirroring the Shanghai part).
+- Effect on existing runs: all 6 runs changed only the line that saves; the probe passes all 6 times; the estimate is no verdict changes.
 
 ## infinite-dup
 
-要求：①每个任务只出现一次；②按顺序；③滚到底能拿到全部任务。判分在每页 1 秒的慢后端下查了 ①②③。隐含：加载中换筛选 / 搜索以后，列表照样能滚到底（Team Tasks 原本的 `reload` 会解锁）。
+Requirements: ① each task appears once; ② in order; ③ scrolling to the end gets every task. The judge checked ①②③ against a slow backend at 1 second per page. Implied: after the filter or the search changes mid-load, the list still scrolls to the end (the original `reload` of Team Tasks releases the lock).
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
 | throttle-1500 | fail | **LEAK** |
 | busy-stuck-after-filter | fail | **LEAK** |
 | dedupe-no-guard | pass | OK |
 
-- **throttle-1500**（LEAK）：用 1.5 秒节流代替「请求在路上就不再发」。每页 1 秒的判分场景挡得住；把判分场景换成每页 2 秒，同一份代码「90 tasks are listed twice」。参考答案、dedupe-no-guard 和 6 局真实产出在 2 秒下都通过（实测）。
-  建议：判分场景的慢页改成 2–2.5 秒（判分循环 40 秒上限够用），或者两档速度各跑一遍。
-- **busy-stuck-after-filter**（隐含要求，LEAK）：加了锁，但 `reload` 不解���、`finally` 只在同一代时解锁。下一页在路上时换一次 Status，锁就永远不放，换回 All 后只剩 20 条（探针实测 20 / 100）。
-  建议：加一步「滚动触发下一页后立刻换 Status 再换回 All，滚到底条数等于总数」。参考答案和 6 局真实产出都过（它们的 `reload` 都会重置锁）。
-- 影响已有局：两个探针在 6 局上全过，猜测不会改判。
+- **throttle-1500** (LEAK): a 1.5-second throttle in place of "no new request while one is on its way". It holds in the judging scenario at 1 second per page; with the scenario at 2 seconds per page the same code gives "90 tasks are listed twice". The reference solution, dedupe-no-guard and the 6 real runs all pass at 2 seconds (measured).
+  Suggested fix: make the slow page 2–2.5 seconds in the judging scenario (the 40-second limit of the judge's loop is enough), or run both speeds.
+- **busy-stuck-after-filter** (implied requirement, LEAK): a lock was added, but `reload` does not release it and the `finally` releases it only within the same generation. Change Status once while the next page is on its way and the lock is never released; back on All only 20 rows are left (the probe measured 20 / 100).
+  Suggested fix: add a step, "after scrolling triggers the next page, change Status at once and change back to All; scrolled to the end, the row count equals the total". The reference solution and the 6 real runs all pass (their `reload` resets the lock).
+- Effect on existing runs: both probes pass on all 6 runs; the estimate is no verdict changes.
 
 ## row-actions-kbd
 
-要求：①键盘能到 Complete / Star；②能用键盘操作；③在它们上面时可见；④其他时候藏着直到悬停；⑤悬停照样显示。判分全都查了，④ 只看焦点所在那一行。
+Requirements: ① the keyboard reaches Complete / Star; ② they can be used from the keyboard; ③ they are visible while focus is on them; ④ otherwise hidden until hover; ⑤ hover shows them as before. The judge checked all of these, ④ only on the focused row.
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
-| focus-opacity-only | pass | **STRICT**（the focused Complete button is covered by something） |
+| focus-opacity-only | pass | **STRICT** (the focused Complete button is covered by something) |
 | all-rows-on-focus | fail | **LEAK** |
-| tab-but-invisible | fail | OK（the focused Complete button is invisible） |
+| tab-but-invisible | fail | OK (the focused Complete button is invisible) |
 | buttons-only-focus | pass | OK |
 
-- **focus-opacity-only**（STRICT）：很常见的正确写法——`.inbox-row:focus-within .row-actions { opacity: 1; }`，`pointer-events` 仍只由 `:hover` 打开。键盘操作不需要 `pointer-events`；鼠标一到行上 `:hover` 就打开了，用户看不出差别。判分的「被遮挡」检查用 `elementFromPoint`，而命中测试会跳过 `pointer-events: none` 的元素，于是判成「被别的东西盖住」。
-  建议：做命中测试前把焦点元素临时设成 `pointer-events: auto`（子元素显式 auto 会覆盖继承的 none），测完恢复；或改用 `elementsFromPoint` 后只看焦点元素前面有没有不透明的、不属于它的元素。
-- **all-rows-on-focus**（LEAK）：`.inbox-list:focus-within .row-actions` 让焦点在列表里任何一处时所有行都显示按钮，违背「其他时候保持隐藏、设计要安静的列表」。
-  建议：Tab 到某行的按钮时，另取一行（不是焦点所在行）检查其 Star 按钮 `shown < 0.2`。
-- 影响已有局：6 局都是 `:hover, :focus-within { opacity: 1; pointer-events: auto; }`，STRICT 的修正不影响；探针 6 次全过，猜测不会改判。
+- **focus-opacity-only** (STRICT): a very common correct approach, `.inbox-row:focus-within .row-actions { opacity: 1; }`, with `pointer-events` still switched on by `:hover` alone. Keyboard use does not need `pointer-events`, and as soon as the mouse is over the row `:hover` switches it on, so the user sees no difference. The judge's "covered" check uses `elementFromPoint`, and hit testing skips elements with `pointer-events: none`, hence the verdict "covered by something".
+  Suggested fix: set the focused element to `pointer-events: auto` for the duration of the hit test (an explicit auto on a child overrides the inherited none) and restore it after; or use `elementsFromPoint` and only look at whether something opaque that does not belong to the focused element lies in front of it.
+- **all-rows-on-focus** (LEAK): `.inbox-list:focus-within .row-actions` shows the buttons on every row while focus is anywhere in the list, against "stay hidden otherwise; the design wants a quiet list".
+  Suggested fix: when Tab reaches a row's button, take another row (not the focused one) and check its Star button has `shown < 0.2`.
+- Effect on existing runs: all 6 runs use `:hover, :focus-within { opacity: 1; pointer-events: auto; }`, which the STRICT fix does not touch; the probe passes all 6 times; the estimate is no verdict changes.
 
 ## draft-keep
 
-要求：①每个任务有自己的草稿；②「在这个浏览器里」保存；③刷新后还在；④回到这个任务（含前进后退）还在；⑤发出后清掉。判分查了 ①③④⑤，② 没查。隐含：发送失败时草稿不该丢（「Once the comment is posted, its draft is cleared」）。
+Requirements: ① each task has its own draft; ② saved "in this browser"; ③ still there after a reload; ④ still there on coming back to the task (including back and forward); ⑤ cleared once posted. The judge checked ①③④⑤ and not ②. Implied: a failed send should not lose the draft ("Once the comment is posted, its draft is cleared").
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
 | session-storage | fail | **LEAK** |
 | clear-before-post | fail | **LEAK** |
-| posted-stays-stored | fail | OK（a posted comment comes back as a draft after a reload） |
+| posted-stays-stored | fail | OK (a posted comment comes back as a draft after a reload) |
 | json-map | pass | OK |
 
-- **session-storage**（LEAK）：localStorage 换成 sessionStorage。同一标签页里刷新、前进后退都在，新开一个标签页草稿就没了。判分全程只用一个标签页。
-  建议：最后加一步在同一个 context 里新开页面打开 `/inbox/4`，草稿 B 还在。这条算不算错有一点余地（「in this browser」也可以读成「这个标签页」），但新标签页拿不到草稿，用户会觉得「丢了」，建议算错。
-- **clear-before-post**（LEAK）：一点 Post 就先删存储里的草稿再发请求，发送失败时刷新就丢。
-  建议：加一个评论 POST 返回 500 的步骤，失败后刷新草稿仍在。属于隐含要求，优先级低。
-- 影响已有局：6 局都用 localStorage，且都在成功后才清；探针 12 次全过，猜测不会改判。
+- **session-storage** (LEAK): sessionStorage instead of localStorage. The draft survives a reload and back / forward in the same tab and is gone in a new tab. The judge uses a single tab throughout.
+  Suggested fix: at the end, open a new page in the same context at `/inbox/4`; draft B is still there. Whether this is wrong has a little room ("in this browser" can also be read as "this tab"), but a new tab without the draft looks like "lost" to the user, so the suggestion is to count it as wrong.
+- **clear-before-post** (LEAK): on Post the stored draft is deleted before the request is sent, so after a failed send a reload loses it.
+  Suggested fix: add a step in which the comment POST returns 500; after the failure and a reload the draft is still there. An implied requirement, low priority.
+- Effect on existing runs: all 6 runs use localStorage and clear only after success; the probes pass all 12 times; the estimate is no verdict changes.
 
 ## attach-upload
 
-要求：①点文件名在任务里（列表下方）显示预览；②再点一次收起；③刚上传的文件直接打开预览；④大于 1 MB 立刻拒绝、不发请求；⑤提示语说出上限；⑥1 MB 以内照常上传。判分查了 ①③④⑤⑥；② 没查；⑥ 只在拒绝之前查了一次。
+Requirements: ① clicking a file name shows a preview inside the task (below the list); ② another click hides it; ③ a file just uploaded opens in the preview directly; ④ a file over 1 MB is refused at once, without a request; ⑤ the message states the limit; ⑥ files up to 1 MB upload as before. The judge checked ①③④⑤⑥; not ②; ⑥ only once, before the refusal.
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
 | no-hide-on-second-click | fail | **LEAK** |
 | stuck-after-refuse | fail | **LEAK** |
-| msg-or-smaller | pass | **STRICT**（picking a 1.5 MB file does not say right away that files can be up to 1 MB） |
-| mirasim-r1-verbatim | pass | **STRICT**（同上） |
+| msg-or-smaller | pass | **STRICT** (picking a 1.5 MB file does not say right away that files can be up to 1 MB) |
+| mirasim-r1-verbatim | pass | **STRICT** (same) |
 | mirasim-r1-reworded | pass | OK |
 
-（变体名里的 `mirasim-r1` 指 Mirasim Computer Use 第 1 局。）
+(`mirasim-r1` in the variant names stands for Mirasim Computer Use run 1.)
 
-- **msg-or-smaller / mirasim-r1-verbatim**（STRICT，已造成误判）：判分要求提示里同时有「1 MB」和 `too (big|large)|up to|limit|max|exceed|larger|bigger|over` 之一。「Files must be 1 MB or smaller.」明确说了上限，却因为 smaller 不在名单里判错；「at most 1 MB」「1 MB or less」也会被判错。mirasim-r1-verbatim 是 Mirasim Computer Use 第 1 局改过的 4 个文件原样，mirasim-r1-reworded 只把这句换成「Files can be up to 1 MB.」——后者通过，说明 **Mirasim Computer Use 第 1 局只栽在措辞上，其余（预览、再点收起、上传后打开、先拒不发请求）都对**。
-  建议：只要求选完 1.5 MB 文件后 1.5 秒内页面上出现含「1 MB」（`(^|[^\d.])1(\.0+)?\s?MB`，已排除「1.5 MB」）的文字，去掉关键词名单；或名单加上 `smaller|less|at most|limit`。
-  影响已有局：Mirasim Computer Use 第 1 局应改判通过（mirasim-r1-reworded 已实测）；其余 5 局的提示都含「maximum」或「limit」，不受影响。
-- **no-hide-on-second-click**（LEAK）：点文件名总是打开，再点不收起。
-  建议：第一次点开后再点一次，iframe 应消失（或不可见）。
-- **stuck-after-refuse**（LEAK）：大小检查放在 `setUploading(true)` 之后，拒绝时直接 return 没复位，按钮一直是「Uploading…」且禁用，之后小文件也传不了。判分把超限放在最后一步；而且 Playwright 的 `setInputFiles` 对禁用的 input 照样能塞文件，只在拒绝后再传一次也抓不到。
-  建议：拒绝后再查 file input 没被禁用、页面上没有「Uploading」，再传一个小文件能存上。
-- 影响已有局：两个探针在 6 局上 12 次全过，猜测不会改判。
+- **msg-or-smaller / mirasim-r1-verbatim** (STRICT, has caused a wrong verdict): the judge requires the message to contain "1 MB" together with one of `too (big|large)|up to|limit|max|exceed|larger|bigger|over`. "Files must be 1 MB or smaller." states the limit plainly and is still judged wrong because smaller is not on the list; "at most 1 MB" and "1 MB or less" would be judged wrong too. mirasim-r1-verbatim is the 4 files Mirasim Computer Use run 1 changed, as they are; mirasim-r1-reworded only rewrites that sentence as "Files can be up to 1 MB.", and it passes. So **Mirasim Computer Use run 1 failed on the wording alone; everything else (preview, hiding on a second click, opening after upload, refusing before any request) is right**.
+  Suggested fix: only require that text containing "1 MB" (`(^|[^\d.])1(\.0+)?\s?MB`, which excludes "1.5 MB") appears on the page within 1.5 seconds of picking the 1.5 MB file, and drop the keyword list; or add `smaller|less|at most|limit` to the list.
+  Effect on existing runs: Mirasim Computer Use run 1 should become a pass (measured with mirasim-r1-reworded); the messages of the other 5 runs all contain "maximum" or "limit" and are not affected.
+- **no-hide-on-second-click** (LEAK): clicking the file name always opens the preview and a second click does not hide it.
+  Suggested fix: after the first click opens it, click once more; the iframe should be gone (or not visible).
+- **stuck-after-refuse** (LEAK): the size check sits after `setUploading(true)` and on a refusal returns without resetting it, so the button stays at "Uploading…", disabled, and small files cannot be uploaded afterwards either. The judge has the oversized file as its last step; and Playwright's `setInputFiles` puts files even into a disabled input, so just uploading again after the refusal would not catch it.
+  Suggested fix: after the refusal check that the file input is not disabled and the page has no "Uploading", then upload a small file and check it is saved.
+- Effect on existing runs: the two probes pass all 12 times on the 6 runs; the estimate is no verdict changes.
 
 ## back-keeps-place
 
-要求（390×844）：①用 ← Inbox 回列表保持位置；②用浏览器后退也是；③列表「完全」停在原处；④已加载的几批都还在；⑤刚看过的任务高亮。判分都查了，但 ③ 的做法是挑视口 45% 处的一行、回来后这行离顶部的距离差 ≤ 48px；⑤ 只看 `li` 自己的类名 / aria 和 `li` 自己的底色、阴影、描边。
+Requirements (390×844): ① going back to the list with ← Inbox keeps the place; ② so does the browser's Back; ③ the list stays "exactly" where it was; ④ the batches already loaded are still there; ⑤ the task just viewed is highlighted. The judge checked all of these, but ③ by picking a row at 45% of the viewport and allowing its distance from the top to differ by ≤ 48px on return; ⑤ only by the `li`'s own class name / aria and the `li`'s own background, shadow and outline.
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
 | center-selected-row | fail | **LEAK** |
-| highlight-on-link | pass | **STRICT**（after the browser's Back the task you came from is not highlighted） |
-| highlight-only | fail | OK（…the task you came from is not in the list） |
+| highlight-on-link | pass | **STRICT** (after the browser's Back the task you came from is not highlighted) |
+| highlight-only | fail | OK (…the task you came from is not in the list) |
 
-- **center-selected-row**（LEAK）：不恢复滚动位置，回来时把刚看过的那行 `scrollIntoView({ block: "center" })`。判分挑的行本来就在视口中段，居中后差不到 48px。探针改为打开靠视口顶部的一行、比较滚动区 `scrollTop`（容差 8px）：参考答案前后一致，本变体 1500 → 1292。
-  建议：比较打开前、回来后滚动区的 `scrollTop`（容差 8px 左右），或挑靠视口顶部的行。题面说「exactly where it was」，严格比较不算过严。
-- **highlight-on-link**（STRICT）：刚看过的行用自己的类 `recently-opened`，底色画在行里的标题链接上（链接占了行的大部分），不设 `aria-current`。肉眼看就是高亮，但 `li` 本身底色透明，判分认不出。
-  建议：`highlighted` 的「看样式」一档同时比较 `li` 和它的直接子元素（标题链接）与邻行对应元素的底色 / 阴影 / 描边；或者截两行的图比较像素。
-- 影响已有局：6 局都用现成的 `selected` 类（ChatGPT Computer Use 第 3 局另把窄屏隐藏改成 `visibility: hidden`、回来时重取已加载的页），探针 6 次全过，猜测不会改判。
+- **center-selected-row** (LEAK): does not restore the scroll position; on return it calls `scrollIntoView({ block: "center" })` on the row just viewed. The row the judge picks is in the middle of the viewport already, so centring moves it by less than 48px. The probe instead opens a row near the top of the viewport and compares the scroll area's `scrollTop` (8px tolerance): the same before and after on the reference solution, 1500 → 1292 on this variant.
+  Suggested fix: compare the scroll area's `scrollTop` before opening and after coming back (a tolerance of about 8px), or pick a row near the top of the viewport. The ticket says "exactly where it was", so a strict comparison is not too strict.
+- **highlight-on-link** (STRICT): the row just viewed gets a class of its own, `recently-opened`, and the background is drawn on the title link inside the row (the link takes up most of the row), without `aria-current`. To the eye it is highlighted, but the `li` itself has a transparent background and the judge does not recognise it.
+  Suggested fix: in the "look at the style" tier of `highlighted`, compare the `li` and its direct child (the title link) with the corresponding elements of a neighbouring row for background / shadow / outline; or screenshot the two rows and compare pixels.
+- Effect on existing runs: all 6 runs use the existing `selected` class (ChatGPT Computer Use run 3 also changes hiding on narrow screens to `visibility: hidden` and refetches the loaded pages on return); the probe passes all 6 times; the estimate is no verdict changes.
 
 ## old-links
 
-要求：①旧链接到任务的新地址；②后退回到打开前的页面；③已删任务的旧链接说清「已删除」；④说出谁删的、何时删的；⑤不存在的旧链接仍说找不到；⑥任务页的 Copy link 复制新地址；⑦收件箱里的；⑧行 ⋯ 菜单里的；⑨新任务的链接能打开。判分查了除 ④ 以外的全部（③ 还要求出现任务名，题面没写，但 TaskGone 本来就带任务名，不算过严）。
+Requirements: ① an old link goes to the task's new address; ② Back returns to the page from before it was opened; ③ an old link to a deleted task says clearly "deleted"; ④ and who deleted it and when; ⑤ an old link to a task that never existed still says not found; ⑥ Copy link on the task page copies the new address; ⑦ so does the one in the Inbox; ⑧ and the one in the row's ⋯ menu; ⑨ a new task's link opens. The judge checked everything except ④ (for ③ it also requires the task's name, which the ticket does not state, but TaskGone carries the name already, so this is not too strict).
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
 | deleted-no-who-when | fail | **LEAK** |
-| removed-wording | pass | **STRICT**（an old link to a deleted task does not say the task was deleted） |
-| push-not-replace | fail | OK（Back … goes to /tasks/2） |
+| removed-wording | pass | **STRICT** (an old link to a deleted task does not say the task was deleted) |
+| push-not-replace | fail | OK (Back … goes to /tasks/2) |
 | redirect-to-tombstone | pass | OK |
 
-- **deleted-no-who-when**（LEAK）：旧链接指向已删任务时自己写了「This task was deleted」+ 任务名，没有谁、何时。题面括号里明确写了「(by whom, and when)」。
-  建议：等到页面上出现删除人（判分场景下是 Alex Kim）和一个日期。探针在参考答案、redirect-to-tombstone、removed-wording 和 12 局真实产出上都过。注意成员列表是异步加载的，「by Alex Kim」比标题晚出现，要等而不是读一次（探针第一版就因为只读一次误判了参考答案）。
-- **removed-wording**（STRICT，有争议）：已删除页写「This task was removed」「… was removed by X on …」，谁、何时都有，只是不用 deleted 这个词。题面要「say clearly that the task was deleted」，用 removed 意思一样，但这次 12 局都是直接用 TaskGone，没人改措辞。
-  建议：`/deleted|removed/i`，同时「不存在的链接」那一步的反向检查也换成同一组词。优先级低。
-- 影响已有局：探针 12 次全过，猜测不会改判。
+- **deleted-no-who-when** (LEAK): for an old link to a deleted task it writes its own "This task was deleted" plus the task's name, with no who or when. The ticket says so in brackets: "(by whom, and when)".
+  Suggested fix: wait until the person who deleted it (Alex Kim in the judging scenario) and a date appear on the page. The probe passes on the reference solution, redirect-to-tombstone, removed-wording and the 12 real runs. Note that the member list loads asynchronously, so "by Alex Kim" appears later than the heading; wait for it instead of reading once (the first version of the probe read once and wrongly failed the reference solution).
+- **removed-wording** (STRICT, debatable): the deleted page says "This task was removed" and "… was removed by X on …"; who and when are both there, only the word deleted is not used. The ticket wants it to "say clearly that the task was deleted"; removed means the same, but all 12 runs this time use TaskGone as it is and nobody changed the wording.
+  Suggested fix: `/deleted|removed/i`, and the reverse check in the "link to a task that never existed" step changes to the same words. Low priority.
+- Effect on existing runs: the probe passes all 12 times; the estimate is no verdict changes.
 
 ---
 
-## 最该先改的 5 处
+## The five things to fix first
 
-1. **attach-upload 提示语的关键词名单**（STRICT，已造成误判）：只认「1 MB」本身，去掉或放宽 `too big|up to|limit|…` 名单。改完重判，Mirasim Computer Use 第 1 局应改判通过。
-2. **row-actions-kbd 的「被遮挡」检查**（STRICT）：`elementFromPoint` 会跳过 `pointer-events: none`，把「focus 只开 opacity」这种很常见的正确写法判成被遮挡。命中测试前临时把焦点元素设成 `pointer-events: auto`。眼下 6 局都没踩到，但换个模型很容易踩。
-3. **assignee-combo 方向键实际没测**（LEAK，题面明写）：加一个能匹配多人的输入，按几次方向键再回车，选中的不是第一个；顺带把一次输入换成姓的一部分（prefix-only），把高亮检测收紧到 `aria-activedescendant` / 焦点 / `aria-selected`（sr-blind）。
-4. **old-links 没查「谁删的、何时删的」**（LEAK，题面明写）：等到删除人和日期出现。
-5. **attach-upload 没查「再点收起」和「拒绝后还能传」**（LEAK，题面明写）：各加一步；拒绝后要查 input 没被禁用，光靠 `setInputFiles` 查不出来。
+1. **The keyword list for the attach-upload message** (STRICT, has caused a wrong verdict): accept "1 MB" itself and drop or loosen the `too big|up to|limit|…` list. On a re-judge after the change, Mirasim Computer Use run 1 should become a pass.
+2. **The "covered" check of row-actions-kbd** (STRICT): `elementFromPoint` skips `pointer-events: none` and judges the very common correct "focus only switches opacity" approach as covered. Set the focused element to `pointer-events: auto` for the hit test. None of the 6 runs hits it so far, but another model easily could.
+3. **The arrow keys of assignee-combo are not really tested** (LEAK, stated in the ticket): add an input that matches several people, press the arrow keys a few times and then Enter, and the pick is not the first; along the way change one input to part of a surname (prefix-only) and tighten highlight detection to `aria-activedescendant` / focus / `aria-selected` (sr-blind).
+4. **old-links does not check "who deleted it and when"** (LEAK, stated in the ticket): wait for the person and the date to appear.
+5. **attach-upload does not check "another click hides it" or "uploads still work after a refusal"** (LEAK, stated in the ticket): add a step for each; after the refusal check the input is not disabled, which `setInputFiles` alone cannot show.
 
-其次：board-dnd 拖回本列不该移动（drop-toggles，绕过了本题的核心 bug 照样全过）、due-off-by-one 西雅图也查编辑页和再存、infinite-dup 的慢页改到 2 秒以上、back-keeps-place 比较 `scrollTop` 并放宽高亮检测。
+Next: in board-dnd a drop back into the same column should not move the card (drop-toggles gets around the task's core bug and still passes everything), due-off-by-one checks the edit page and a second save in Seattle too, the slow page of infinite-dup goes above 2 seconds, back-keeps-place compares `scrollTop` and loosens highlight detection.
 
 ---
 
-## 修订结果（按「全部改」）
+## What was changed (as decided: "change them all")
 
-9 题全部改了判分，只动 `tasks/controls.ts`（back-keeps-place 另加了一个 sanity 变体），一题一个提交。每题都跑了：`harness/selftest.ts`（种子判失败、参考答案判通过）、`harness/audit.ts`（全部 OK），再用改后的判分对这题**全部**已有局只读重判（用只读的重判脚本，不写回各局的判分结果）。
+The judges of all 9 tasks were changed, touching only `tasks/controls.ts` (back-keeps-place also got one sanity variant), one commit per task. For every task these were run: `harness/selftest.ts` (the seed is judged fail, the reference solution pass), `harness/audit.ts` (all OK), and a read-only re-judge of **all** existing runs of the task with the changed judge (with the read-only re-judge script, which does not write back to the runs' verdicts).
 
-| 题 | 收紧 | 放宽 | 自检 | 变体 | 只读重判 |
+| Task | Stricter | Looser | Self-test | Variants | Read-only re-judge |
 |---|---|---|---|---|---|
-| board-dnd | 开头拖回本列松手不能保存、不能离开 To do；最后让 Done 卡（任务 9）保存失败，拖到 To do 后要回到 Done、接口不变 | — | 1/1 | 4/4 | 8 局不变 |
-| assignee-combo | 加一步输入 "a"（匹配多人）、↓↓↑↓ 后高亮不能还在第一项，回车选中的就是高亮的；找 Diego 改输入 "alvar"；高亮只认 aria-activedescendant / 焦点 / 唯一 aria-selected，去掉只看类名那档 | — | 1/1 | 4/4 | 8 局不变 |
-| due-off-by-one | 西雅图也重开编辑页查回显，再只改 Notes 存一次查日期不变 | — | 1/1 | 3/3 | 6 局不变 |
-| infinite-dup | 判分场景慢页 1 秒 → 2.2 秒；滚到底后切 Done、触发下一页时立刻切回 All，再滚到底照查条数、去重、顺序 | — | 1/1 | 3/3 | 6 局不变 |
-| row-actions-kbd | Tab 到一行的 Complete 时，另一行的 Star 要仍隐藏 | 「被遮挡」检查测之前临时给焦点按钮 `pointer-events: auto` | 1/1 | 4/4 | 6 局不变 |
-| draft-keep | 同一浏览器新开标签页打开任务 4，草稿要在；评论 POST 返回 500 时点 Post 后刷新，草稿要还在 | — | 1/1 | 4/4 | 6 局不变 |
-| attach-upload | 打开预览后再点文件名要收起；拒绝过大文件后上传框不能被禁用 / 停在「Uploading…」，再传小文件要能存上 | 超限提示只要新出现「1 MB」（或 1024 KB），不再要求 too big / up to / limit 等词 | 1/1 | 5/5 | **Mirasim Computer Use 第 1 局未通过 → 通过**，其余 5 局不变 |
-| back-keeps-place | 挑视口 20% 处的行（原 45%），位置容差 48px → 16px | 高亮也比较行的直接子元素（标题链接），比较前把鼠标移开排除悬停底色；新增变体 no-highlight（只修位置不高亮，expect fail）确认不误认 | 1/1 | 4/4 | 6 局不变 |
-| old-links | 已删除提示要等到删除人（取自后端记录）和一个日期出现 | 「deleted」或「removed」都算（按已定）；不存在链接那步的反向检查同样认两个词 | 1/1 | 4/4 | 12 局不变 |
-
-- 没有改 expect 的变体：按已定，old-links/removed-wording 本来就是 expect pass，谁删、何时删它也都有。
-- attach-upload 的 Mirasim Computer Use 第 1 局逐局看过：预览、再点收起、上传后打开、超限先拒不发请求都做了，提示是「Files must be 1 MB or smaller.」，改后判通过符合题面。没有写回这一局的判分结果，留给正式重判。
-- infinite-dup 改后判一局要 40–70 秒（原来 30 秒上下），因为多了一次「换筛选后滚到底」。
-- 9 题改完以后又整组跑了一遍自检和变体，结果见下一行。
-  整组复跑：自检 9/9 OK，变体 35/35 OK（34 个原有 + no-highlight）。
+| board-dnd | At the start, a drop back into the same column must not save and must not leave To do; at the end, a Done card (task 9) whose save fails must return to Done after a drag to To do, with the API unchanged | — | 1/1 | 4/4 | 8 runs unchanged |
+| assignee-combo | An added step types "a" (matching several people); after ↓↓↑↓ the highlight must not still be on the first option, and Enter picks the highlighted one; Diego is now found by typing "alvar"; the highlight is only accepted as aria-activedescendant / focus / the only aria-selected, without the class-name tier | — | 1/1 | 4/4 | 8 runs unchanged |
+| due-off-by-one | In Seattle too, the edit page is reopened and what it shows is checked, then only Notes is changed and saved and the date must not change | — | 1/1 | 3/3 | 6 runs unchanged |
+| infinite-dup | The slow page of the judging scenario goes from 1 second to 2.2 seconds; after scrolling to the end, switch to Done and, as the next page is triggered, straight back to All, then scroll to the end again and check count, duplicates and order as before | — | 1/1 | 3/3 | 6 runs unchanged |
+| row-actions-kbd | When Tab reaches a row's Complete, another row's Star must still be hidden | Before the "covered" check the focused button is given `pointer-events: auto` for the test | 1/1 | 4/4 | 6 runs unchanged |
+| draft-keep | Task 4 opened in a new tab of the same browser must have the draft; when the comment POST returns 500, the draft must still be there after clicking Post and reloading | — | 1/1 | 4/4 | 6 runs unchanged |
+| attach-upload | With the preview open, another click on the file name must hide it; after an oversized file is refused the upload field must not be disabled or stuck at "Uploading…", and a small file must then save | The size-limit message only has to newly show "1 MB" (or 1024 KB), without words such as too big / up to / limit | 1/1 | 5/5 | **Mirasim Computer Use run 1 not passed → pass**, the other 5 runs unchanged |
+| back-keeps-place | Picks a row at 20% of the viewport (was 45%); the position tolerance goes from 48px to 16px | The highlight also compares the row's direct child (the title link), with the mouse moved away first to rule out the hover background; a new variant no-highlight (fixes the position without highlighting, expect fail) confirms nothing is recognised by mistake | 1/1 | 4/4 | 6 runs unchanged |
+| old-links | The deleted-task notice must come to show the person who deleted it (taken from the backend record) and a date | "deleted" or "removed" both count (as decided); the reverse check in the step for a link to a task that never existed accepts both words too | 1/1 | 4/4 | 12 runs unchanged |
+
+- No variant's expect was changed: as decided, old-links/removed-wording was expect pass from the start, and it does say who deleted the task and when.
+- Mirasim Computer Use run 1 of attach-upload was looked at in detail: the preview, hiding on a second click, opening after upload and refusing an oversized file before any request are all done, and the message is "Files must be 1 MB or smaller."; passing it after the change agrees with the ticket. The run's verdict was not written back; that is left to the formal re-judge.
+- After the change one infinite-dup run takes 40–70 seconds to judge (about 30 seconds before), because of the extra "scroll to the end after a filter change".
+- After all 9 tasks were changed, the self-test and the variants were run again for the whole group; the result is on the next line.
+  Rerun of the whole group: self-test 9/9 OK, variants 35/35 OK (the 34 existing ones + no-highlight).

docs/judge-audits/flows.md

@@ -1,214 +1,214 @@
-# 判分审计：流程题（10 题）
+# Judge audit: flow tasks (10 tasks)
 
-本文只保留 v1 题单里的题。
+Only tasks in the v1 suite are kept here.
 
-变体在 `audit-variants/<题 id>/`，共 45 个（33 个 expect fail、12 个 expect pass）。全部用 `npx tsx harness/audit.ts -j 1 <题id>` 跑过，没有 ERR。
-真实产出：7 道修 bug 题逐局看了 diff；3 道长任务另写了一个只读探针（复制一份代码后在模拟后端（apps/mock-api/server.mjs）+ Vite 上走判分没走的场景），ChatGPT Computer Use、Mirasim Computer Use 两侧的局全部跑过一遍。
+The variants are in `audit-variants/<id>/`, 45 in all (33 expect fail, 12 expect pass). All were run with `npx tsx harness/audit.ts -j 1 <id>`, with no ERR.
+Real runs: for the 7 bug-fix tasks the diff of every run was read; for the 3 long tasks a read-only probe was written as well (it copies the code and then, on the mock backend (apps/mock-api/server.mjs) + Vite, goes through the scenarios the judge does not), and it was run over every run on both sides, ChatGPT Computer Use and Mirasim Computer Use.
 
-## 总表
+## Summary
 
-| 题 id | 要求条数 | 没被检查的要求 | 变体数 | LEAK | STRICT | 真实产出里的误判 |
+| Task id | Requirements | Requirements not checked | Variants | LEAK | STRICT | Wrong verdicts on real runs |
 |---|---|---|---|---|---|---|
-| login-return | 2 | 无 | 4 | 0 | 0 | 无（6 局） |
-| viewer-role | 7 | 设置页、收件箱评论框、附件上传、编辑页；被拒后恢复原状实际上没检查到（竞态） | 5 | 1 | 1 | 无（6 局都是先等服务器再删） |
-| session-expiry | 5 | 任务页 Save description 过期后重发 | 4 | 0 | 1 | 无（6 局） |
-| checkout-return | 7 | 不会再从旧收银台链接付款（只靠后端）；假报「浏览器拦截」 | 4 | 1 | 0 | 无（8 局） |
-| balance-refresh | 4 | 无 | 3 | 0 | 0 | 无（8 局） |
-| white-screen | 3 | 已经存成 null 的任务能重新打开；清空截止日期后显示的内容 | 4 | 2 | 0 | 无（8 局改法完全相同） |
-| wrong-param | 3 | Active 往下滚的后续页（小） | 3 | 0 | 0 | 无（6 局） |
-| activity-log | 19 | Orders / Attachments 选项能不能用；单个任务上叠 Person；Newer / Older 撤掉没有；Load more 时换筛选回到第一批 | 6 | 3 | 0 | 无（24 局探针全过） |
-| notify-center | 16 | 页面上的 Mark all as read；未读醒目 / 已读变暗；「离开页面」后不再留着 | 5 | 2 | 0 | 无（17 局探针全过，Mirasim Computer Use 有一局试跑本来就判失败） |
-| recurring | 17 | 任务页 / 编辑页取消再勾上；在已完成的一条上 Stop repeating；This and future 改系列里其他未完成的；每天任务晚完成；Cancel 按钮 | 7 | 4 | 1 | 无（10 局探针全过） |
-| **合计** | | | **45** | **13** | **3** | **无** |
-
-结论先说：这些题没有找到已有局判错，通过率高看起来是模型真的做对了（修 bug 题的改法几乎都一模一样，长任务探针也全过）。但判分有 13 处放得过错的实现，其中 viewer-role 那处是写法上的漏洞（核心要求实际上没检查到），长任务里有几处是前几轮的要求在后面轮次只检查了一半。
+| login-return | 2 | none | 4 | 0 | 0 | none (6 runs) |
+| viewer-role | 7 | The Settings page, the Inbox comment box, attachment upload, the edit page; putting things back after a refusal is in fact not checked (a race) | 5 | 1 | 1 | none (all 6 runs wait for the server before removing) |
+| session-expiry | 5 | Resending Save description on the task page after the session expires | 4 | 0 | 1 | none (6 runs) |
+| checkout-return | 7 | No further payment from the old checkout link (left to the backend alone); a false "blocked by the browser" | 4 | 1 | 0 | none (8 runs) |
+| balance-refresh | 4 | none | 3 | 0 | 0 | none (8 runs) |
+| white-screen | 3 | A task already saved as null can be opened again; what is shown after the due date is cleared | 4 | 2 | 0 | none (the 8 runs make exactly the same change) |
+| wrong-param | 3 | Later pages when scrolling down Active (minor) | 3 | 0 | 0 | none (6 runs) |
+| activity-log | 19 | Whether the Orders / Attachments options work; Person stacked on a single task; whether Newer / Older were removed; going back to the first batch when the filter changes after Load more | 6 | 3 | 0 | none (the probe passes on all 24 runs) |
+| notify-center | 16 | Mark all as read on the page; unread standing out / read dimmed; no longer kept after "leaving the page" | 5 | 2 | 0 | none (the probe passes on all 17 runs; one Mirasim Computer Use trial run was judged fail to begin with) |
+| recurring | 17 | Unticking and ticking again on the task page / edit page; Stop repeating on a completed task; This and future changing the other open tasks of the series; a daily task completed late; the Cancel button | 7 | 4 | 1 | none (the probe passes on all 10 runs) |
+| **Total** | | | **45** | **13** | **3** | **none** |
+
+The conclusion first: no existing run of these tasks was found to be judged wrongly, and the high pass rate looks like the model really getting them right (the changes in the bug-fix tasks are nearly all identical, and the long-task probe passes everywhere). But the judges let wrong implementations through in 13 places. The one in viewer-role is a hole in how the judge is written (the core requirement is in fact not checked), and several in the long tasks are requirements from earlier rounds that are only half checked in later rounds.
 
 ---
 
 ## login-return
 
-要求：① 从 `/tasks/4` 这类链接登录后回到该任务；② 从 `/inbox/5?status=active` 登录后回到收件箱那条并带上筛选。另外判分查了「已登录时开 `/login?next=` 直接去 next」和「next 不能带出站」，都不是工单写的，但种子的注释点名了 open redirect，基线 Team Tasks 本来就这样，不算越界。
+Requirements: ① after signing in from a link such as `/tasks/4`, return to that task; ② after signing in from `/inbox/5?status=active`, return to that Inbox item with the filter. The judge also checked "opening `/login?next=` while signed in goes straight to next" and "next must not lead off the site"; the ticket states neither, but the seed's comment names the open redirect and the baseline Team Tasks already behaves this way, so this is not overreach.
 
-变体：`drop-query`、`tasks-only`、`open-redirect`（fail）都判失败；`url-origin`（pass，用 `new URL` 比同源）判通过。没有问题。
+Variants: `drop-query`, `tasks-only` and `open-redirect` (fail) are all judged fail; `url-origin` (pass, compares origins with `new URL`) is judged pass. No problems.
 
 ## viewer-role
 
-要求：只读成员 ① 任务列表没有新建框；② 列表没有 Edit / Delete / 勾选；③ 任务页没有改动控件；④ 看板没有；⑤「anywhere」（收件箱、设置、编辑页等）；服务器拒绝时 ⑥ 要告诉用户、⑦ 恢复原状。
+Requirements: for a read-only member ① the task list has no new-task box; ② the list has no Edit / Delete / checkbox; ③ the task page has no controls that change anything; ④ nor has the Board; ⑤ "anywhere" (Inbox, Settings, the edit page and so on); when the server refuses, ⑥ tell the user and ⑦ put things back as they were.
 
-没检查的：⑤ 只看了收件箱的按钮和 ⋯ 菜单，设置页的 Invite / Remove、收件箱的评论框和附件上传、`/tasks/:id/edit` 都没看（种子只破坏了一行角色比较，修根因就全好，风险低）。⑥⑦ 只在任务列表删除和勾选上查。
+Not checked: for ⑤ only the Inbox's buttons and ⋯ menu were looked at; Invite / Remove on the Settings page, the Inbox comment box and attachment upload, and `/tasks/:id/edit` were not (the seed breaks a single line that compares roles, so fixing the root cause fixes everything; low risk). ⑥⑦ are only checked on deleting and ticking in the task list.
 
-**LEAK `toast-no-rollback`**（判通过）：乐观删除、被拒时弹错误提示、但不把任务放回来。原因不是漏写，是**竞态**：判分 `page.goto(url)` 后一看到 Delete 按钮就切成 Viewer 并点下去，而 React StrictMode 开发模式下首屏的 `GET /api/tasks` 发了两次，第二次的响应落在乐观删除之后，把任务又放回列表（服务器没删，响应里自然有它）。我在变体目录里加日志确认：立刻点，任务先 9 条再变回 10 条；先等 1.5 秒再点，任务消失、不再回来。所以「after the server refused the delete, the task vanished」这一步在现在的写法下几乎触发不了，种子本身（`role-only`）也是栽在「没提示」而不是「消失」上。
-- 建议：`page.goto(url)` 后 `await settle(page)`（`harness/env.ts` 里已有）或等 1 秒再切 Viewer、再点 Delete；「消失」改成在 1.5 秒内反复看，只要有一次不见就判失败。
-- 对已有局（猜测）：6 局 `deleteTask` 都改回了「先等服务器再拿掉」，改判分不影响结论。
+**LEAK `toast-no-rollback`** (judged pass): an optimistic delete that shows an error notice on refusal but does not put the task back. The cause is not something left out of the judge, it is a **race**: after `page.goto(url)` the judge switches to Viewer and clicks as soon as it sees the Delete button, while under React StrictMode in development the first-screen `GET /api/tasks` is sent twice; the second response lands after the optimistic delete and puts the task back in the list (the server did not delete it, so the response naturally has it). Logging added in the variant's directory confirmed this: clicked at once, the tasks go to 9 and then back to 10; clicked after waiting 1.5 seconds, the task disappears and does not come back. So as written, the step "after the server refused the delete, the task vanished" can hardly ever trigger, and the seed itself (`role-only`) fails on "no notice", not on "vanished".
+- Suggested fix: after `page.goto(url)`, `await settle(page)` (already in `harness/env.ts`) or wait 1 second before switching to Viewer and clicking Delete; for "vanished", look repeatedly over 1.5 seconds and fail if the task is missing even once.
+- Existing runs (estimate): the `deleteTask` of all 6 runs went back to "wait for the server, then remove", so changing the judge changes no verdict.
 
-**STRICT `alert-refusal`**（判失败：the user is not told）：被拒时用原生 `window.alert` 说「You don't have permission…」，任务留在原处。判分只数页面 innerText 里的拒绝字样，看不到原生弹框；`ctx.dialogs` 其实记下了它。
-- 建议：「告诉了用户」= 页面文字多了拒绝字样，或 `ctx.dialogs` 多了一条 alert 且消息匹配同一正则。
-- 对已有局（猜测）：没有局用 alert，不影响。
+**STRICT `alert-refusal`** (judged fail: the user is not told): on refusal a native `window.alert` says "You don't have permission…" and the task stays where it is. The judge only counts refusal wording in the page's innerText and cannot see the native dialog; `ctx.dialogs` did in fact record it.
+- Suggested fix: "the user was told" = the page text gained refusal wording, or `ctx.dialogs` gained an alert whose message matches the same regex.
+- Existing runs (estimate): no run uses alert; not affected.
 
-其余：`role-only`、`silent-rollback` 判失败，`optimistic-rollback`（pass，乐观删除 + 失败回滚并抛错）判通过。
+The rest: `role-only` and `silent-rollback` are judged fail; `optimistic-rollback` (pass: optimistic delete, rollback on failure, error thrown) is judged pass.
 
 ## session-expiry
 
-要求：① 过期时原地要求重新登录；② 保留已输入的；③ 登录后完成按下的保存；④ 任务页 Save description 也一样；⑤ 收件箱发评论也一样。
+Requirements: ① on expiry, ask for sign-in again in place; ② keep what was typed; ③ after sign-in, finish the save that was pressed; ④ the same for Save description on the task page; ⑤ the same for posting a comment in the Inbox.
 
-没检查的：④。种子是全局的（把 401 处理换成退回登录页），按全局修法 ④ 自然好；只有「每个页面各自接 401」的修法会漏，现实中没见到。
+Not checked: ④. The seed is global (it replaces the 401 handling with going back to the sign-in page), and a global fix makes ④ right as a matter of course; only a fix that "handles 401 on each page separately" would miss it, which has not been seen in practice.
 
-**STRICT `own-error-text`**（判失败：a wrong password in the sign-in prompt shows no error）：重新登录对话框密码错时显示自己的文案「That password didn't work…」。工单没提密码错要怎么提示，判分却要求文字匹配 `/wrong|incorrect|invalid|didn.?t match|try again/`。
-- 建议：放宽为「出现任何提示（`role=alert`、或页面多出一段文字）且对话框还在」，或者只要求「错密码时没有保存、对话框还在」，不看措辞。
-- 对已有局（猜测）：改过对话框的 ChatGPT Computer Use 第 2、3 局和 Mirasim Computer Use 第 1 局都保留了服务器的错误文字，判通过的不受影响。
+**STRICT `own-error-text`** (judged fail: a wrong password in the sign-in prompt shows no error): on a wrong password the re-sign-in dialog shows a text of its own, "That password didn't work…". The ticket says nothing about how a wrong password should be reported, yet the judge requires text matching `/wrong|incorrect|invalid|didn.?t match|try again/`.
+- Suggested fix: loosen to "some notice appears (`role=alert`, or a new piece of text on the page) and the dialog is still there", or only require "on a wrong password nothing is saved and the dialog is still there", regardless of wording.
+- Existing runs (estimate): ChatGPT Computer Use runs 2 and 3 and Mirasim Computer Use run 1, which changed the dialog, all keep the server's error text; runs judged pass are not affected.
 
-其余：`no-retry`、`patch-only`（只给 PATCH 接重新登录，发评论直接报错）判失败；`editable-email-dialog`（pass）判通过。
+The rest: `no-retry` and `patch-only` (re-sign-in wired only to PATCH; posting a comment just errors) are judged fail; `editable-email-dialog` (pass) is judged pass.
 
 ## checkout-return
 
-要求：① 收银台在新标签页打开；② Credits 页留在原处；③「Cancel and return」回来要作废订单；④ 回来的那页显示 Cancelled；⑤ 留着的 Credits 页也显示；⑥ 不能再从收银台链接付款；⑦ 照常付款。
+Requirements: ① the checkout opens in a new tab; ② the Credits page stays where it is; ③ coming back with "Cancel and return" must void the order; ④ the page returned to shows Cancelled; ⑤ so does the Credits page that stayed open; ⑥ no further payment from the checkout link; ⑦ paying works as before.
 
-没检查的：⑥ 只间接靠「后端状态是 cancelled」保证（后端拒绝给已作废的订单付款），可以接受。
+Not checked: ⑥ is only ensured indirectly through "the backend state is cancelled" (the backend refuses payment for a voided order), which is acceptable.
 
-**LEAK `noopener-says-blocked`**（判通过）：`window.open(url, "_blank", "noopener")` 返回 null，收银台照常打开，Credits 页却同时显示「Your browser blocked the checkout tab」。这是把原有提示弄错，工单没写，严重度低。
-- 建议（可选）：点 Buy 后断言页面上没有 `/blocked/i`。
-- 对已有局（猜测）：真实局都是先开 `about:blank` 再 `opener = null`，返回值非空，不受影响。
+**LEAK `noopener-says-blocked`** (judged pass): `window.open(url, "_blank", "noopener")` returns null; the checkout opens as usual, yet the Credits page shows "Your browser blocked the checkout tab" at the same time. This breaks an existing notice, which the ticket does not mention; low severity.
+- Suggested fix (optional): after clicking Buy, assert the page has no `/blocked/i`.
+- Existing runs (estimate): the real runs all open `about:blank` first and then set `opener = null`; the return value is not null, so they are not affected.
 
-其余：`newtab-only`、`cancel-local-only` 判失败；`blank-first-canceled`（pass，先开空白页再跳、状态写成 Canceled）判通过。
+The rest: `newtab-only` and `cancel-local-only` are judged fail; `blank-first-canceled` (pass: opens a blank page first and then navigates, status written as Canceled) is judged pass.
 
 ## balance-refresh
 
-要求：① 余额自己更新；② 订单显示 Paid；③ 铃铛收到 Payment received；④ 付款在切回来之后才完成也要更新。全部检查到。另外查了「断线一次后照样更新」，工单没写，但 EventSource 断线会自动重连，修对事件的实现都会过，不算过严。
+Requirements: ① the balance updates by itself; ② the order shows Paid; ③ the bell gets Payment received; ④ it also updates when the payment completes only after switching back. All checked. The judge also checks "it still updates after one dropped connection"; the ticket does not state it, but EventSource reconnects by itself after a drop, so any implementation that fixes the events passes, and this is not too strict.
 
-变体：`focus-only`（只接回聚焦刷新）、`poll-billing`（Credits 页轮询，铃铛不动）判失败；`sse-only`（pass，只修事件，不接回聚焦刷新）判通过。没有问题。
+Variants: `focus-only` (only refreshes on regaining focus) and `poll-billing` (the Credits page polls, the bell does not move) are judged fail; `sse-only` (pass: only the events fixed, no refresh on focus) is judged pass. No problems.
 
 ## white-screen
 
-要求：① 在编辑页保存后任务页不再白屏；② 已经白屏的任务（「from then on that task always opens white」）能重新打开；③ 其他任务照常。
+Requirements: ① after a save on the edit page the task page is no longer blank; ② a task that already went blank ("from then on that task always opens white") can be opened again; ③ other tasks as before.
 
-没检查的：② 判分从干净数据开始，没有「已经存成 null 的任务」；保存后只看标题有没有出来，不看页面显示的内容和保存的值。
+Not checked: ② the judge starts from clean data, with no "task already saved as null"; after the save it only looks at whether the title shows, not at what the page displays or what was saved.
 
-**LEAK `omit-empty-on-save`**（判通过）：只改保存端，截止日期为空时不发 `dueDate`。白屏不再出现，但清空截止日期再保存不生效（任务页仍显示 Due Dec 1, 2026），已经存成 null 的任务照样白屏。
-**LEAK `empty-string-due`**（判通过）：只改保存端，空截止日期存成 `""`。`formatDay("")` 不抛错，任务页显示「Due Invalid Date」。
-- 建议：判分开头先用后端把某条任务的 `dueDate` 设成 null（`PATCH /tasks/3 {dueDate:null}`），再打开它，要求不白屏；清空截止日期那一步之后要求后端 `dueDate` 为空（null / 不存在 / ""），且任务页不出现 `Dec 1, 2026` 和 `Invalid Date`。
-- 对已有局（猜测）：8 局都是在详情页把条件改成 `task.dueDate &&`，加了这些检查仍全过。
+**LEAK `omit-empty-on-save`** (judged pass): only the saving side is changed, and `dueDate` is not sent when the due date is empty. The blank screen no longer appears, but clearing the due date and saving has no effect (the task page still shows Due Dec 1, 2026), and a task already saved as null still opens blank.
+**LEAK `empty-string-due`** (judged pass): only the saving side is changed, and an empty due date is saved as `""`. `formatDay("")` does not throw, and the task page shows "Due Invalid Date".
+- Suggested fix: at the start of the judge, set the `dueDate` of some task to null through the backend (`PATCH /tasks/3 {dueDate:null}`), then open it and require that it is not blank; after the step that clears the due date, require the backend's `dueDate` to be empty (null / absent / "") and the task page to show neither `Dec 1, 2026` nor `Invalid Date`.
+- Existing runs (estimate): all 8 runs change the condition on the detail page to `task.dueDate &&`; they still all pass with these checks added.
 
-其余：`undefined-check`（换成 `!== undefined`）判失败；`no-due-text`（pass，没截止日期时写 No due date）判通过。
+The rest: `undefined-check` (changed to `!== undefined`) is judged fail; `no-due-text` (pass: writes No due date when there is none) is judged pass.
 
 ## wrong-param
 
-要求：① 选 Done 只列已完成；② 选 Active 只列未完成；③ 标题旁的数量跟着变。都查到（Done 还往下滚到底）。Active 只看第一页和数量，后续页没滚，影响很小。
+Requirements: ① choosing Done lists only completed tasks; ② choosing Active lists only open ones; ③ the count next to the heading follows. All checked (Done is also scrolled to the end). For Active only the first page and the count are looked at, and later pages are not scrolled; the effect is small.
 
-变体：`client-filter`（前端过滤、数量按过滤后算）、`loadmore-drops-status` 判失败；`send-both`（pass）判通过。没有问题。
+Variants: `client-filter` (filters in the front end, the count taken after filtering) and `loadmore-drops-status` are judged fail; `send-both` (pass) is judged pass. No problems.
 
-## activity-log（5 轮）
+## activity-log (5 rounds)
 
-按最终代码应满足的要求（第 4 轮把 Newer / Older 换成 Load more，其余各轮的要求都保留）：
-- 第 1 轮：Type 筛选含 All activity / Tasks / Comments / Attachments / Orders ①；Person 筛选 ②；筛选在地址里，刷新、分享同样 ③；Back 撤销上一次筛选 ④；换筛选从第一批开始 ⑤。
-- 第 2 轮：任务页 View activity 链接 ⑥；说明是哪个任务 ⑦；能回到全部 ⑧；Type 和 Person 能叠在上面 ⑨。
-- 第 3 轮：每条有 Delete ⑩；原生确认框写明哪一条 ⑪；取消不删 ⑫；删后筛选不变 ⑬；Viewer 没有 Delete ⑭。
-- 第 4 轮：Load more 往下加 20 条 ⑮；没有更多时消失 ⑯；Newer / Older 撤掉 ⑰（地址不再需要页码，非要求）。
-- 第 5 轮：删一条后 Load more 不漏 ⑱；有新动静后不重复 ⑲。
+Requirements the final code should meet (round 4 replaces Newer / Older with Load more; the requirements of the other rounds all stay):
+- Round 1: a Type filter with All activity / Tasks / Comments / Attachments / Orders ①; a Person filter ②; filters are in the address, the same on reload and when shared ③; Back undoes the last filter ④; a filter change starts from the first batch ⑤.
+- Round 2: a View activity link on the task page ⑥; it says which task ⑦; a way back to all activity ⑧; Type and Person can be stacked on it ⑨.
+- Round 3: every entry has Delete ⑩; a native confirm that names the entry ⑪; cancelling deletes nothing ⑫; the filter is unchanged after a delete ⑬; a Viewer has no Delete ⑭.
+- Round 4: Load more adds 20 entries below ⑮; it goes away when there are no more ⑯; Newer / Older are removed ⑰ (the address no longer needs a page number; not a requirement).
+- Round 5: after deleting an entry, Load more skips nothing ⑱; after new activity it repeats nothing ⑲.
 
-判分查到了 ②③④⑥⑦⑧⑩–⑯⑱⑲，① 只查了 Tasks 和 Comments，⑨ 只查了 Type。没查到的：
+The judge checked ②③④⑥⑦⑧⑩–⑯⑱⑲; ① only for Tasks and Comments, ⑨ only for Type. Not checked:
 
-**LEAK `newer-older-kept`**（判通过）：加了 Load more，第 1 轮的 Newer / Older 按钮还留在下面。第 4 轮的改需求只检查了「有 Load more」，没检查「旧的分页没了」。
-- 建议：Load more 那一段加一句：`main` 里没有名字匹配 `/^(← )?newer|older( →)?$/i` 的按钮或链接（也可以用 `/page \d+ of \d+/i` 文字辅助）。
-- 对已有局（猜测）：24 局源码里都没有 Newer / Older 按钮（两局试跑里只有 `listOlder` 这个函数名），不受影响。
+**LEAK `newer-older-kept`** (judged pass): Load more was added and the Newer / Older buttons of round 1 are still underneath. For the changed requirement of round 4 the judge only checks "there is a Load more", not "the old pagination is gone".
+- Suggested fix: add to the Load more part: `main` has no button or link whose name matches `/^(← )?newer|older( →)?$/i` (the text `/page \d+ of \d+/i` can help too).
+- Existing runs (estimate): the source of the 24 runs has no Newer / Older button (two trial runs only have a function named `listOlder`); not affected.
 
-**LEAK `orders-type-wrong`**（判通过）：Orders 选项取值写成 `orders`，选 Orders 永远是空列表。
-- 建议：把 Type 依次选 Orders（初始数据里有 2 条 order.*）对一下；Attachments 初始数据里是 0 条，要查就先用后端传一个附件。
-- 对已有局：探针 24 局全部 Orders 显示 2 条、与后端一致，不受影响（实测）。
+**LEAK `orders-type-wrong`** (judged pass): the value of the Orders option is written `orders`, and choosing Orders always gives an empty list.
+- Suggested fix: choose Orders for Type and compare (the initial data has 2 order.* entries); Attachments has 0 entries in the initial data, so to check it first upload an attachment through the backend.
+- Existing runs: on all 24 runs the probe shows 2 entries for Orders, matching the backend; not affected (measured).
 
-**LEAK `person-ignored-on-task`**（判通过）：单个任务的视图里忽略 Person。
-- 建议：在「one task's activity + Type Comments」之后，再选 Person = Priya Shah，与 `activityOf(api, {taskId: 2, actorId: "m2", type: "comment"})` 对上。
-- 对已有局：探针 24 局「任务 2 + Priya Shah」都显示 3 条、与后端一致，不受影响（实测）。
+**LEAK `person-ignored-on-task`** (judged pass): the single-task view ignores Person.
+- Suggested fix: after "one task's activity + Type Comments", also choose Person = Priya Shah and compare with `activityOf(api, {taskId: 2, actorId: "m2", type: "comment"})`.
+- Existing runs: on all 24 runs the probe shows 3 entries for "task 2 + Priya Shah", matching the backend; not affected (measured).
 
-⑤（Load more 之后换筛选回到第一批）没写变体：`matches` 允许显示条数在 n 到全部之间，多显示了也算对；这一条用户几乎看不出差别，不建议专门加。
+No variant was written for ⑤ (going back to the first batch when the filter changes after Load more): `matches` allows the number of entries shown to be anywhere from n to all of them, so showing more counts as right; a user can hardly tell the difference, and adding a check just for this is not advised.
 
-**「回到全部」的放宽有没有太松**：之前定过，题面有两种读法时两种都算对，所以「回到全部」既可以只清任务、也可以连 Type 一起清。`link-clears-all`（pass，`<Link to="/activity">` 连 Type 一起清）判通过，符合这条原则。我又推了几种错的回法：「Show all」不清任务只清 Type、或回去后列表不刷新，放宽后的分支都会因为「既不是全部评论、也不是全部活动」而判失败；回到全部后还重选一次 Comments 核对，所以放宽没有把错的放过去。不算太松。
+**Is the loosening of "back to all activity" too loose?** It was decided earlier that when a ticket can be read two ways, both count as right, so "back to all" may clear only the task or the Type along with it. `link-clears-all` (pass: `<Link to="/activity">` clears the Type too) is judged pass, in line with that rule. A few wrong ways of going back were also worked through: a "Show all" that keeps the task and clears only the Type, or a list that does not refresh after going back; the loosened branch fails both, as "neither all comments nor all activity". After going back to all, Comments is also selected once more and compared, so the loosening lets nothing wrong through. Not too loose.
 
-其余：`no-dedupe`（第 5 轮 bug）、`filter-replace-history`（Back 撤不回）判失败。
+The rest: `no-dedupe` (the round 5 bug) and `filter-replace-history` (Back cannot undo) are judged fail.
 
-## notify-center（5 轮）
+## notify-center (5 rounds)
 
-最终要求：
-- 第 1 轮：侧栏链接带未读数 ①；`/notifications` 全部、最新在前 ②；未读醒目 ③；每条 Mark as read ④；Mark all as read ⑤；点关于任务的通知打开任务并标已读 ⑥；铃铛、侧栏、页面一致且实时 ⑦。
-- 第 2 轮：All / Unread 标签 ⑧；标签在地址里，刷新 / Back 保持 ⑨；Unread 里标为已读的留着（变暗）直到离开页面或刷新 ⑩。
-- 第 3 轮：弹出提示带标题和 View ⑪；在通知页上不弹 ⑫。
-- 第 4 轮（改需求）：只弹指派给我的 ⑬；标签页标题 `(3) Team Tasks` ⑭。
-- 第 5 轮：断线时铃铛面板和通知页显示 Reconnecting… ⑮；连回来补上 ⑯。
+Final requirements:
+- Round 1: a sidebar link with the unread count ①; `/notifications` lists everything, newest first ②; unread ones stand out ③; Mark as read on each ④; Mark all as read ⑤; clicking a notification about a task opens the task and marks it read ⑥; bell, sidebar and page agree and are live ⑦.
+- Round 2: All / Unread tabs ⑧; the tab is in the address and survives reload / Back ⑨; in Unread, an item marked read stays (dimmed) until you leave the page or reload ⑩.
+- Round 3: a popup with the title and View ⑪; no popup while on the notifications page ⑫.
+- Round 4 (changed requirement): popups only for tasks assigned to me ⑬; the tab title `(3) Team Tasks` ⑭.
+- Round 5: while disconnected, the bell panel and the notifications page show Reconnecting… ⑮; what was missed is caught up on reconnect ⑯.
 
-判分查到了 ①②④⑥–⑨⑪–⑯，⑩ 只查了「刷新后不留」。第 3 轮的「弹出」被第 4 轮收窄，判分按第 4 轮查（`popup-any-task` 判失败），跟上了。没查到的：
+The judge checked ①②④⑥–⑨⑪–⑯; for ⑩ only "not kept after a reload". Round 3's "popup" is narrowed by round 4, and the judge follows round 4 (`popup-any-task` is judged fail), so it keeps up. Not checked:
 
-**LEAK `no-mark-all-on-page`**（判通过）：通知页上没有 Mark all as read（铃铛面板里原本有一个）。工单第 1 轮写在页面要求里。
-- 建议：通知页上要求有 `/mark all as read/i` 按钮；可在最后「断线补上」之前点一下，核对三处都变 0、后端全部已读。
-- 对已有局：17 局页面上都有这个按钮（实测）。
+**LEAK `no-mark-all-on-page`** (judged pass): the notifications page has no Mark all as read (the bell panel had one already). Round 1 of the ticket lists it among the page's requirements.
+- Suggested fix: require a `/mark all as read/i` button on the notifications page; it can be clicked just before the final "catch up after a disconnect" to check all three places go to 0 and everything is read in the backend.
+- Existing runs: all 17 runs have this button on the page (measured).
 
-**LEAK `kept-survives-leave`**（判通过）：Unread 里「标为已读先留着」的记录放在模块变量里，刷新会清，离开页面（点进任务）再 Back 回来，那条已读的还在 Unread 里。⑩ 写的是「until you leave the page or reload」，判分只查了 reload。
-- 建议：在 Unread 里标一条已读后，点开另一条通知进任务页，再 Back 回 Unread，要求那条不在了（或按之前「留着但显示成已读也算对」的读法显示成已读——但「离开过页面」后两种读法都应该不在）。
-- 对已有局：17 局（Mirasim Computer Use 那局本来判失败的试跑、探针点不到，除外）离开再回来都不留，不受影响（实测）。
+**LEAK `kept-survives-leave`** (judged pass): the record of "marked read but kept for now" in Unread lives in a module variable. A reload clears it; after leaving the page (clicking through to a task) and coming Back, the read item is still in Unread. ⑩ says "until you leave the page or reload", and the judge only checks the reload.
+- Suggested fix: after marking one item read in Unread, open another notification to go to its task page, come Back to Unread and require the item to be gone (or, under the earlier reading "kept but shown as read is right too", shown as read; but once the page "has been left" it should be gone under both readings).
+- Existing runs: none of the 17 runs keeps it after leaving and coming back (except the one Mirasim Computer Use trial run that was judged fail to begin with, where the probe cannot click through); not affected (measured).
 
-③「未读醒目 / 已读变暗」没写变体：属于样式，判分不好通用地查。可以退一步只查「同一页上未读和已读的那两条 class 或 font-weight / opacity 不同」，可选。
+No variant was written for ③ "unread stands out / read is dimmed": it is styling, hard for a judge to check in a general way. A fallback would be to check only that "an unread and a read item on the same page differ in class or in font-weight / opacity"; optional.
 
-**「留着但显示成已读」的放宽有没有太松**：放宽只允许「A 在 Unread 里但显示成已读（Mark as read 不可见或禁用）」。完全没过滤的实现仍会因为「一开始就已读的 seeded[2] 还在」判失败；把 A 当未读留着（按钮还能点）的也判失败。`link-tabs`（pass，链接式标签、`?filter=unread`、留着的显示 Read 字样）判通过。放宽本身不松，松的是上面那条「离开页面」没查，和这处放宽无关。
+**Is the loosening of "kept but shown as read" too loose?** It only allows "A is in Unread but shown as read (Mark as read not visible or disabled)". An implementation that does not filter at all still fails, because "seeded[2], read from the start, is still there"; one that keeps A as unread (the button still clickable) fails too. `link-tabs` (pass: link-style tabs, `?filter=unread`, kept items showing the word Read) is judged pass. The loosening itself is not loose; what is loose is the missing "leaving the page" check above, which has nothing to do with it.
 
-其余：`no-catchup` 判失败。
+The rest: `no-catchup` is judged fail.
 
-## recurring（5 轮）
+## recurring (5 rounds)
 
-最终要求：
-- 第 1 轮：四个入口完成都生成下一条 ①；复制标题 / 优先级 / 描述 / 负责人 / 规则 ②；未完成 ③；按规则从旧截止日往后（每 2 周 +14 天、每月同日）④（被第 4 轮部分改掉）；没截止日期从今天算 ⑤；完成的那条保持完成 ⑥。
-- 第 2 轮：任务页、看板卡片显示规则和截止日期 ⑦⑧。
-- 第 3 轮：保存时问 Only this task / This and future tasks ⑨；只改这一条，下一条按旧系列 ⑩；这条及以后：这一条、系列里每条未完成的、之后生成的都改 ⑪；取消或 Escape 不保存、保留输入 ⑫。
-- 第 4 轮（改需求）：每天 / 每周晚完成从完成日算 ⑬；每月仍从旧截止日加一个月 ⑭。
-- 第 5 轮：取消再��上不再生成 ⑮；Stop repeating 关掉这一条和系列里未完成的，不删任何东西 ⑯⑰。
+Final requirements:
+- Round 1: completing through any of the four entry points creates the next task ①; title / priority / description / assignee / rule are copied ②; it is not done ③; the date moves on from the old due date by the rule (every 2 weeks +14 days, monthly on the same day) ④ (partly changed by round 4); with no due date, counted from today ⑤; the completed one stays completed ⑥.
+- Round 2: the task page and the Board card show the rule and the due date ⑦⑧.
+- Round 3: saving asks Only this task / This and future tasks ⑨; only this one changes and the next follows the old series ⑩; this and future: this one, every open one in the series and those created later all change ⑪; Cancel or Escape saves nothing and keeps the input ⑫.
+- Round 4 (changed requirement): a daily / weekly task completed late counts from the completion day ⑬; monthly still adds a month to the old due date ⑭.
+- Round 5: unticking and ticking again creates nothing more ⑮; Stop repeating switches off this task and the open ones of the series, deleting nothing ⑯⑰.
 
-判分对改需求跟得上：`r1-rule-kept`（每周晚完成仍从旧截止日算）、`monthly-from-completion`（每月也按完成日算）都判失败。但第 3、4、5 轮的要求都只用了「最好走」的那条路径检查：
+The judge keeps up with the changed requirements: `r1-rule-kept` (a weekly task completed late still counted from the old due date) and `monthly-from-completion` (monthly also counted from the completion day) are both judged fail. But the requirements of rounds 3, 4 and 5 are each checked along the "easiest" path only:
 
-**LEAK `retick-guard-list-only`**（判通过）：「不再生成」只在 toggleDone（列表 / 看板）里防了，任务页 Completed 框和编辑页 Status 走 updateTask，取消再勾上又多一条。判分只在任务列表上取消再勾。
-**LEAK `stop-only-this`**（判通过）：Stop repeating 只关当前这一条。判分是在系列里唯一未完成的那条上点 Stop，所以「连同系列里未完成的」从来没被检查。
-**LEAK `future-skips-open-siblings`**（判通过）：This and future tasks 不改系列里其他未完成的任务。判分做这一步时系列里只有一条未完成的。
-**LEAK `daily-late-from-due`**（判通过）：第 4 轮只改了每周，每天的任务晚完成仍从旧截止日算。判分里唯一的每天任务没有截止日期。
-- 建议（四条可以一起加到现有流程里，我的探针就是这么走的）：
-  1. 任务页上把一条已完成的重复任务取消再勾上，未完成的同名任务仍只有 1 条；
-  2. 在已完成的上一条（如 n2）的任务页上点 Stop repeating，要求下一条（n3）的 `recurrence` 变空；若认为已完成的任务不该有 Stop 按钮，也可以改成「系列里同时有两条未完成时（取消勾选上一条），在其中一条上点 Stop，另一条也停」；
-  3. 同样的「两条未完成」状态下对其中一条选 This and future tasks 改优先级，另一条也要变；
-  4. 加一条每天、截止日期在过去的任务，完成后下一条是明天。
-- 对已有局：探针 10 局四项全过（实测），加上后不影响结论。
+**LEAK `retick-guard-list-only`** (judged pass): "creates nothing more" is guarded only in toggleDone (list / Board); the Completed box on the task page and Status on the edit page go through updateTask, and unticking and ticking again adds another task. The judge only unticks and ticks in the task list.
+**LEAK `stop-only-this`** (judged pass): Stop repeating only switches off the current task. The judge clicks Stop on the only open task of the series, so "along with the open ones of the series" was never checked.
+**LEAK `future-skips-open-siblings`** (judged pass): This and future tasks does not change the other open tasks of the series. When the judge does this step the series has a single open task.
+**LEAK `daily-late-from-due`** (judged pass): round 4 was only applied to weekly; a daily task completed late still counts from the old due date. The only daily task in the judge has no due date.
+- Suggested fixes (all four can go into the existing flow together, which is the route the probe takes):
+  1. On the task page, untick and tick again a completed repeating task; there is still just 1 open task of that name;
+  2. On the task page of the completed previous task (such as n2) click Stop repeating and require the `recurrence` of the next one (n3) to become empty; if a completed task is thought not to need a Stop button, this can instead be "with two open tasks in the series (untick the previous one), click Stop on one of them and the other stops too";
+  3. In the same "two open tasks" state, choose This and future tasks on one of them to change the priority; the other must change too;
+  4. Add a daily task with a due date in the past; after completing it the next one is due tomorrow.
+- Existing runs: the probe passes all four on the 10 runs (measured); adding them changes no verdict.
 
-**STRICT `scope-radio-dialog`**（判失败：does not ask Only this task / This and future tasks）：问范围用两个单选项加 OK / Cancel（Google 日历就是这样）。判分只认名字匹配 `/only this/i` 的**按钮**。工单只说「ask whether … Only this task or This and future tasks」，单选 + 确认是合理写法。
-- 建议：认 `getByRole("button")` 或 `getByRole("radio")` 或 `getByLabel`；是单选时先选中，再点对话框里的 OK / Save / Confirm / Apply。
-- 对已有局（猜测）：10 局都用两个按钮，不受影响。
+**STRICT `scope-radio-dialog`** (judged fail: does not ask Only this task / This and future tasks): the scope is asked with two radio options plus OK / Cancel (as Google Calendar does). The judge only accepts a **button** whose name matches `/only this/i`. The ticket only says "ask whether … Only this task or This and future tasks"; radio plus confirm is a reasonable way to do it.
+- Suggested fix: accept `getByRole("button")` or `getByRole("radio")` or `getByLabel`; for a radio, select it first and then click OK / Save / Confirm / Apply in the dialog.
+- Existing runs (estimate): all 10 runs use two buttons; not affected.
 
-⑫ 只查了 Escape，没查 Cancel 按钮，风险小，未写变体。
+For ⑫ only Escape was checked, not the Cancel button; the risk is small and no variant was written.
 
 ---
 
-## 最该先改的 5 处
+## The five things to fix first
 
-1. **viewer-role：去掉「点得太早」的竞态**。`goto` 后先 `settle(page)` 再切 Viewer、点 Delete；顺带把原生 alert 也算作「告诉了用户」。现在「被拒后任务消失」这一步实际上查不到，乐观删除不回滚的实现会判通过。
-2. **recurring：第 3、5 轮的系列操作补一条「系列里有两条」的路径**（任务页取消再勾上、在已完成的一条上 Stop、两条未完成时 This and future），再加一条每天晚完成。四个 LEAK 都是前几轮要求只查了最好走的路径。
-3. **activity-log：第 4 轮改需求后检查 Newer / Older 已撤掉；Orders 选项和「单个任务 + Person」各对一次数据**。
-4. **white-screen：先放一条 dueDate 为 null 的任务再打开；清空截止日期后核对后端值和页面不出现旧日期 / Invalid Date**。只改保存端的半截修法现在会判通过。
-5. **放宽三处只认一种写法的检查**：recurring 的范围对话框认单选；session-expiry 的错密码提示不看措辞；notify-center 再补「页面上的 Mark all as read」和「离开页面后 Unread 不再留着已读的」。
+1. **viewer-role: remove the "clicked too early" race.** After `goto`, `settle(page)` first, then switch to Viewer and click Delete; along the way count a native alert as "told the user". At present the step "the task vanishes after the refusal" in fact cannot be checked, and an optimistic delete without rollback is judged pass.
+2. **recurring: give the series operations of rounds 3 and 5 a path with "two tasks in the series"** (unticking and ticking again on the task page, Stop on a completed task, This and future with two open tasks), plus a daily task completed late. All four LEAKs are requirements from earlier rounds checked along the easiest path only.
+3. **activity-log: after the changed requirement of round 4, check that Newer / Older are gone; compare the data once each for the Orders option and for "single task + Person".**
+4. **white-screen: put in a task whose dueDate is null and open it; after clearing the due date, check the backend value and that the page shows neither the old date nor Invalid Date.** A half fix on the saving side alone is judged pass at present.
+5. **Loosen three checks that accept a single way of doing it**: the scope dialog of recurring accepts radios; the wrong-password notice of session-expiry is not judged by wording; notify-center also gets "Mark all as read on the page" and "after leaving the page Unread no longer keeps the read item".
 
-按目前的真实产出，这 5 处改完后已有局的结论预计都不变（1、4、5 是猜测；2、3 和 notify-center 两条用探针在全部局上实测过）。
+Going by the real runs so far, the verdicts of existing runs are expected not to change after these 5 fixes (1, 4 and 5 are estimates; 2, 3 and the two notify-center points were measured with the probe on every run).
 
 ---
 
-## 修订结果（第二步）
+## What was changed (second step)
 
-按上面的报告把 13 处 LEAK、3 处 STRICT 全部改进 `tasks/flows.ts`，一题一次改动。自检用 `harness/selftest.ts -j 1`，变体用 `harness/audit.ts -j 1`；抽查重判在真实局的副本上跑（复制代码后判，不写回原来的判分结果）。
+Following the report above, all 13 LEAKs and 3 STRICTs were fixed in `tasks/flows.ts`, one change per task. The self-test is `harness/selftest.ts -j 1`, the variants `harness/audit.ts -j 1`; the spot re-judge runs on copies of real runs (the code is copied and judged, without writing back to the original verdicts).
 
-| 题 | 改了什么 | 自检 | 变体 | 抽查重判 |
+| Task | What changed | Self-test | Variants | Spot re-judge |
 |---|---|---|---|---|
-| viewer-role | 收紧：点 Delete 前先 `settle`，修掉首屏重复请求把任务放回来的竞态；放宽：被拒时原生 alert 说明也算告诉了用户 | OK | 5/5 OK | ChatGPT Computer Use 第 1 局、Mirasim Computer Use 第 2、3 局：通过 → 通过 |
-| session-expiry | 放宽：密码错的提示不看措辞（role=alert 或多出一行字，且还在要密码）；修竞态：两处让会话过期前先 `settle`；新增变体 `no-error-shown` | OK | 5/5 OK | ChatGPT Computer Use 第 2 局（重判 4 次）、第 3 局，Mirasim Computer Use 第 1、2 局：通过 → 通过 |
-| checkout-return | 收紧：收银台打开后 Credits 页不能显示 blocked | OK | 4/4 OK | ChatGPT Computer Use 第 2 局、Mirasim Computer Use 的两局试跑：通过 → 通过 |
-| white-screen | 收紧：先把任务 3 的 dueDate 设成 null 再打开；清空截止日期后后端要清空、页面不显示旧日期；不出现 Invalid Date / NaN；新增变体 `omit-empty-and-guard`、`empty-string-null-guard` | OK | 6/6 OK | ChatGPT Computer Use 第 1 局和一局试跑、Mirasim Computer Use 第 3 局：通过 → 通过 |
-| activity-log | 收紧：Load more 后不能还有 Newer / Older；单个任务 + Type Comments 上再叠 Person；全部活动上选 Orders 对数据。「回到全部」的放宽不变 | OK | 6/6 OK | ChatGPT Computer Use 第 1 局和一局试跑、Mirasim Computer Use 第 4 局和一局试跑：通过 → 通过 |
-| notify-center | 收紧：Unread 里标为已读后离开页面再 Back 回来，那条不能还在；页面上的 Mark all as read 要能用。「留着但显示成已读」的放宽不变 | OK | 5/5 OK | ChatGPT Computer Use 第 3 局和一局试跑、Mirasim Computer Use 第 1 局和一局试跑：通过 → 通过；Mirasim Computer Use 另一局试跑：未通过 → 未通过（同一原因） |
-| recurring | 收紧：每天任务晚完成（任务 10）；任务页 Completed 框取消再勾上；系列里两条未完成时 This and future 和 Stop repeating 都要改到另一条（等到再判）；新建任务数 6 → 7。放宽：问范围认单选加确认按钮 | OK | 7/7 OK | ChatGPT Computer Use 第 1–5 局、Mirasim Computer Use 第 1–5 局共 10 局：都通过 → 通过 |
-| login-return、balance-refresh、wrong-param | 没改：审计没发现 LEAK / STRICT | OK | 10/10 OK | — |
-
-自检 10/10 OK，变体 48/48 OK（原 45 个 + 新增 3 个）。
-
-抽查重判里没有局的结论变化。过程中有两次新判分自己出的误判，已经修好：
-- **session-expiry，ChatGPT Computer Use 第 2 局**：第一次只改了措辞时，这局重判 4 次有 2 次判失败（登录后没完成保存）。原因是判分在页面加载的请求还没回来时就让会话过期，晚到的请求先撞上 401、先弹出重新登录框，挡住了随后的 Save 点击（点击失败被 `.catch` 吞掉）。旧判分也有这个竞态，只是没碰上。加 `settle` 后 4 次都通过。
-- **recurring，ChatGPT Computer Use 第 4 局**：Stop repeating 后马上读另一条任务，而这局是先改这一条、再一条一条改系列里的其他任务，判分读早了。改成等到再判后通过。
+| viewer-role | Stricter: `settle` before clicking Delete, fixing the race in which the duplicate first-screen request puts the task back; looser: a native alert that explains the refusal counts as telling the user | OK | 5/5 OK | ChatGPT Computer Use run 1, Mirasim Computer Use runs 2 and 3: pass → pass |
+| session-expiry | Looser: the wrong-password notice is not judged by wording (role=alert or a new line of text, with the password still being asked for); race fixed: `settle` in two places before the session is made to expire; new variant `no-error-shown` | OK | 5/5 OK | ChatGPT Computer Use run 2 (re-judged 4 times) and run 3, Mirasim Computer Use runs 1 and 2: pass → pass |
+| checkout-return | Stricter: once the checkout has opened, the Credits page must not show blocked | OK | 4/4 OK | ChatGPT Computer Use run 2, the two Mirasim Computer Use trial runs: pass → pass |
+| white-screen | Stricter: the dueDate of task 3 is set to null first and the task then opened; after the due date is cleared the backend must be cleared and the page must not show the old date; no Invalid Date / NaN; new variants `omit-empty-and-guard`, `empty-string-null-guard` | OK | 6/6 OK | ChatGPT Computer Use run 1 and a trial run, Mirasim Computer Use run 3: pass → pass |
+| activity-log | Stricter: no Newer / Older may remain after Load more; Person is stacked on single task + Type Comments; Orders is chosen on all activity and compared with the data. The loosening of "back to all activity" is unchanged | OK | 6/6 OK | ChatGPT Computer Use run 1 and a trial run, Mirasim Computer Use run 4 and a trial run: pass → pass |
+| notify-center | Stricter: after marking an item read in Unread, leaving the page and coming Back, the item must not still be there; Mark all as read on the page must work. The loosening of "kept but shown as read" is unchanged | OK | 5/5 OK | ChatGPT Computer Use run 3 and a trial run, Mirasim Computer Use run 1 and a trial run: pass → pass; another Mirasim Computer Use trial run: not passed → not passed (same cause) |
+| recurring | Stricter: a daily task completed late (task 10); unticking and ticking the Completed box on the task page; with two open tasks in the series, This and future and Stop repeating must both reach the other one (the judge waits for it); new tasks counted go from 6 to 7. Looser: radios plus a confirm button are accepted for the scope question | OK | 7/7 OK | ChatGPT Computer Use runs 1–5 and Mirasim Computer Use runs 1–5, 10 runs: all pass → pass |
+| login-return, balance-refresh, wrong-param | Not changed: the audit found no LEAK / STRICT | OK | 10/10 OK | — |
+
+Self-test 10/10 OK, variants 48/48 OK (the 45 original ones + 3 new).
+
+No run changed verdict in the spot re-judge. Twice along the way the new judge itself gave a wrong verdict; both are fixed:
+- **session-expiry, ChatGPT Computer Use run 2**: when only the wording had been changed at first, this run was judged fail in 2 of 4 re-judges (the save was not finished after sign-in). The cause: the judge let the session expire while the page's loading requests were still out; a late request hit the 401 first and brought up the re-sign-in dialog early, which blocked the click on Save that followed (the failed click was swallowed by `.catch`). The old judge had this race too and just never met it. With `settle` added, all 4 pass.
+- **recurring, ChatGPT Computer Use run 4**: the judge read the other task straight after Stop repeating, while this run changes this task first and then the others of the series one by one; the judge read too early. Changed to wait for it, and the run passes.

docs/judge-audits/interaction-layout.md

@@ -1,245 +1,239 @@
-# 判分审计：交互与布局题（11 题）
+# Judge audit: interaction and layout tasks (11 tasks)
 
-本文只保留 v1 题单里的题。
+Only tasks in the v1 suite are kept here.
 
-变体在 `audit-variants/<题id>/`，共 53 个（含 hover-ghost 原有 3 个）。最后一次全量运行：
+The variants are in `audit-variants/<id>/`, 53 in all (including the 3 that hover-ghost already had). The last full run:
 `npx tsx harness/audit.ts -j 1 divider-lock card-stretch loading-veil sticky-save hover-ghost menu-clip narrow-nav dark-mode scroll-tail mention-trim type-scale`
-→ 37/53 OK，**LEAK 11，STRICT 5**，ERR 0。
+→ 37/53 OK, **LEAK 11, STRICT 5**, ERR 0.
 
-真实产出：11 题 ChatGPT Computer Use、Mirasim Computer Use 两侧全部局的 diff 都看过（每题 6–13 局，全部判通过）；对判分没量到的几处，
-另用只读脚本在已有局的代码上量了 type-scale 8 局、narrow-nav 10 局、dark-mode 6 局（见各节）。
-**没有发现「判了通过但其实没做到题面要求」的局。**
+Real runs: the diffs of every run on both sides, ChatGPT Computer Use and Mirasim Computer Use, were read for all 11 tasks (6–13 runs per task, all judged pass). For a few things the judge did not measure, a read-only script measured them on the code of existing runs: 8 type-scale runs, 10 narrow-nav runs and 6 dark-mode runs (see each section).
+**No run was found that was judged pass without doing what the ticket asks.**
 
-## 总表
+## Summary
 
-| 题 id | 要求条数 | 没被检查的要求 | 变体数 | LEAK | STRICT | 真实产出里的误判 |
+| Task id | Requirements | Requirements not checked | Variants | LEAK | STRICT | Wrong verdicts on real runs |
 |---|---|---|---|---|---|---|
-| divider-lock | 3 | — | 4 | 0 | 0 | 无 |
-| card-stretch | 4 | — | 4 | 0 | 0 | 无 |
-| loading-veil | 5 | 首次加载时的转圈 | 5 | 1 | 1（边缘） | 无 |
-| sticky-save | 4 | Notes 链接（页面已在底部，没法验，可不管） | 4 | 0 | 0 | 无 |
-| hover-ghost | 3 | — | 5 | 0 | **2** | 无 |
-| menu-clip | 3 | — | 4 | 0 | 0 | 无 |
-| narrow-nav | 5 | 宽屏页头不多出 Menu 按钮；按钮在页头里 | 6 | 1 | 1（边缘） | 无 |
-| dark-mode | 4 | 「下次来访」仍记得；次要文字可读 | 5 | 2 | 0 | 无 |
-| scroll-tail | 2 | — | 4 | 0 | 0 | 无 |
-| mention-trim | 6 | 头像还在；↑ 键 | 6 | 2 | 1 | 无 |
-| type-scale | 约 25 | 头像缩写、任务页头 Edit/Copy link、列表页头与筛选不动；Done=12、Comments=14 只查了「在三档里」 | 6 | 5 | 0 | 无 |
-| 合计 |  |  | 53 | 11 | 5 |  |
-
-「改判分会不会影响已有局」一栏里，标「已量」的是用脚本在已有局代码上实测过的，标「猜测」的是只看了 diff。
+| divider-lock | 3 | — | 4 | 0 | 0 | none |
+| card-stretch | 4 | — | 4 | 0 | 0 | none |
+| loading-veil | 5 | The spinner on the first load | 5 | 1 | 1 (borderline) | none |
+| sticky-save | 4 | The Notes link (the page is already at the bottom, so it cannot be tested; can be left) | 4 | 0 | 0 | none |
+| hover-ghost | 3 | — | 5 | 0 | **2** | none |
+| menu-clip | 3 | — | 4 | 0 | 0 | none |
+| narrow-nav | 5 | No extra Menu button in the header on wide screens; the button being in the header | 6 | 1 | 1 (borderline) | none |
+| dark-mode | 4 | Remembered "on a later visit"; secondary text readable | 5 | 2 | 0 | none |
+| scroll-tail | 2 | — | 4 | 0 | 0 | none |
+| mention-trim | 6 | The avatar still being there; the ↑ key | 6 | 2 | 1 | none |
+| type-scale | about 25 | Avatar initials, Edit / Copy link in the task page header, and the list header and filters left alone; Done=12 and Comments=14 only checked as "one of the three sizes" | 6 | 5 | 0 | none |
+| Total |  |  | 53 | 11 | 5 |  |
+
+Under "would changing the judge affect existing runs", "measured" means measured by script on the code of existing runs, and "estimate" means only the diff was read.
 
 ---
 
 ## divider-lock
 
-要求：拖完后页面能点（任务、按钮）；光标恢复；拖动仍能改宽、松手就停。全部查到。
+Requirements: after a drag the page can be clicked (tasks, buttons); the cursor is restored; dragging still changes the width and stops on release. All checked.
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
-| up-on-divider（松手挂在分隔条 onPointerUp，没 capture） | fail | OK |
-| shield-passthrough（挡板 pointer-events: none、去光标，拖动不结束） | fail | OK |
-| pointer-capture（setPointerCapture + onPointerUp） | pass | OK |
-| buttons-released（pointermove 里发现键已松开就结束） | pass | OK |
+| up-on-divider (release handled in the divider's onPointerUp, no capture) | fail | OK |
+| shield-passthrough (the shield gets pointer-events: none and loses the cursor; the drag never ends) | fail | OK |
+| pointer-capture (setPointerCapture + onPointerUp) | pass | OK |
+| buttons-released (ends the drag when pointermove finds the button released) | pass | OK |
 
-10 局全是 mouseup → pointerup，判分无问题。
+All 10 runs change mouseup → pointerup; no problem with the judge.
 
 ## card-stretch
 
-要求：★、Mark done、✕ 能点且不跳转；卡片其他地方（含底栏空白、标题）仍打开任务。全部查到，三个按钮分在三张卡上逐个点，底栏空白处也点了。
+Requirements: ★, Mark done and ✕ can be clicked without navigating; the rest of the card (including the empty part of the footer and the title) still opens the task. All checked; the three buttons are clicked one by one on three different cards, and the empty part of the footer is clicked too.
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
-| raise-footer（整个底栏抬到拉伸层上面） | fail | OK（底栏空白不再打开任务） |
-| after-no-pointer（::after 加 pointer-events: none） | fail | OK |
-| star-only（只修星标） | fail | OK |
-| js-card-click（去掉 ::after，li 上 onClick 跳转） | pass | OK |
+| raise-footer (the whole footer raised above the stretched layer) | fail | OK (the empty part of the footer no longer opens the task) |
+| after-no-pointer (pointer-events: none on ::after) | fail | OK |
+| star-only (only the star fixed) | fail | OK |
+| js-card-click (::after removed, onClick on the li navigates) | pass | OK |
 
-10 局都是 `.card-actions button { position: relative; z-index: 2 }`，判分无问题。
+All 10 runs use `.card-actions button { position: relative; z-index: 2 }`; no problem with the judge.
 
 ## loading-veil
 
-要求：慢网络加载完列表能点开任务；悬停出 ✓/☆；⋯ 能开；切 Status 后同样；加载时继续显示转圈。
+Requirements: on a slow network, once the list has loaded a task can be opened by clicking; hover brings out ✓/☆; ⋯ opens; the same after switching Status; a spinner keeps showing while loading.
 
-没被检查：**首次加载时的转圈**。判分开头只等加载提示消失，从不要求它出现过；只在切筛选那次要求出现。
+Not checked: **the spinner on the first load**. At its start the judge only waits for the loading indicator to go away and never requires it to have appeared; it requires it to appear only on the filter switch.
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
 | raise-links | fail | OK |
 | no-spinner | fail | OK |
-| **no-initial-spinner**（列表里已有任务时才盖遮罩，首次加载什么都没有） | fail | **LEAK** |
+| **no-initial-spinner** (the veil only covers the list when it already has tasks; nothing on the first load) | fail | **LEAK** |
 | visibility-hidden | pass | OK |
-| **fade-with-spinner**（淡出的遮罩加 pointer-events: none，转圈跟着一起淡出、留在 DOM 里） | pass | **STRICT**：the loading indicator never goes away |
+| **fade-with-spinner** (the fading veil gets pointer-events: none; the spinner fades with it and stays in the DOM) | pass | **STRICT**: the loading indicator never goes away |
 
-- LEAK no-initial-spinner：建议 `openInbox` 之前（或换成先 goto 再等）在 1s 内要求列表上出现加载提示（判分场景 1200ms 延迟，300ms 后遮罩出现，时间够）。影响已有局：猜测无——12 局都没给遮罩加「有数据才显示」之类的条件。
-- STRICT fade-with-spinner（边缘）：`loadingShown` 用 Playwright `isVisible()`，它不看 opacity，opacity 0 的转圈算「还在」。这种写法看着点着都对，但 role=status 的「Loading tasks」一直在无障碍树里，读屏会一直读到，算不算对可以商量。若要放宽：`loadingShown` 改用 `checkVisibility({ opacityProperty: true, visibilityProperty: true })`。影响已有局：无（没有一局保留转圈）。
+- LEAK no-initial-spinner: suggested fix, before `openInbox` (or with a goto first and a wait after) require a loading indicator on the list within 1s (the judging scenario has a 1200ms delay and the veil appears after 300ms, so there is time). Effect on existing runs: none by estimate; none of the 12 runs puts a condition such as "only when there is data" on the veil.
+- STRICT fade-with-spinner (borderline): `loadingShown` uses Playwright's `isVisible()`, which does not look at opacity, so a spinner at opacity 0 counts as "still there". This approach looks right and clicks right, but the role=status "Loading tasks" stays in the accessibility tree and a screen reader keeps reading it, so whether it is right is open to discussion. To loosen it: have `loadingShown` use `checkVisibility({ opacityProperty: true, visibilityProperty: true })`. Effect on existing runs: none (no run keeps the spinner).
 
 ## sticky-save
 
-要求：从底部点 Basics，Title 不被工具栏盖、能点；Planning 头一行同样；（其余分节链接同样）；工具栏仍吸顶。判分查了 Basics、Planning、Repeat 和吸顶、保存。Notes 没查（从底部跳 Notes 页面不动，查不出，可不管）。
+Requirements: clicking Basics from the bottom leaves Title not covered by the toolbar and clickable; the same for the first line of Planning; (the same for the other section links); the toolbar still sticks to the top. The judge checked Basics, Planning, Repeat, the sticking and saving. Notes was not checked (from the bottom, jumping to Notes does not move the page, so nothing can be seen; can be left).
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
-| header-only-margin（只扣页头高度） | fail | OK |
-| basics-only（只修 Basics） | fail | OK |
-| unstick-toolbar（去掉吸顶） | fail | OK |
-| scroll-padding-html（html 上 scroll-padding-top） | pass | OK |
+| header-only-margin (only the header's height is allowed for) | fail | OK |
+| basics-only (only Basics fixed) | fail | OK |
+| unstick-toolbar (the sticking removed) | fail | OK |
+| scroll-padding-html (scroll-padding-top on html) | pass | OK |
 
-10 局都是量工具栏高度写进变量再 scroll-margin-top，判分无问题。
+All 10 runs measure the toolbar's height into a variable and use scroll-margin-top; no problem with the judge.
 
 ## hover-ghost
 
-要求：点任务名能打开；悬停出的 ✓/☆ 照常能点；平时不显示。全部查到。
+Requirements: clicking a task's name opens it; the ✓/☆ that hover brings out can be clicked as before; they do not show otherwise. All checked.
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
-| always-visible（样例） | fail | OK |
-| no-restore-pointer（样例） | fail | OK |
-| raise-title-link（链接抬到悬停层上面，按钮被盖） | fail | OK |
-| **display-none-hover**（样例：平时 display: none，悬停 display: flex） | pass | **STRICT**：check threw: locator.evaluate: Timeout 5000ms exceeded |
-| **visibility-hidden**（平时 visibility: hidden，悬停 visible） | pass | **STRICT**：同上 |
+| always-visible (sample) | fail | OK |
+| no-restore-pointer (sample) | fail | OK |
+| raise-title-link (the link raised above the hover layer, covering the buttons) | fail | OK |
+| **display-none-hover** (sample: display: none normally, display: flex on hover) | pass | **STRICT**: check threw: locator.evaluate: Timeout 5000ms exceeded |
+| **visibility-hidden** (visibility: hidden normally, visible on hover) | pass | **STRICT**: same |
 
-- STRICT 原因：判分第一步用 `page.getByRole("button", { name: "Star Release checklist" })` 找按钮，再 `evaluate(checkVisibility)` 看它是否隐藏。`getByRole` 默认不匹配 display: none / visibility: hidden 的元素（它们不在无障碍树里），于是定位不到、等 5 秒抛错。这两种是最常见的隐藏写法，都被判失败。
-- 建议：这一步改成 `getByRole("button", { name, includeHidden: true })`，或者「找不到」直接当作「没显示」；后面悬停之后再找的那几处不用改。
-- 影响已有局：无。13 局都是 opacity + pointer-events（已量：diff 全看过）。
+- Cause of the STRICT: as its first step the judge finds the button with `page.getByRole("button", { name: "Star Release checklist" })` and then uses `evaluate(checkVisibility)` to see whether it is hidden. By default `getByRole` does not match elements with display: none / visibility: hidden (they are not in the accessibility tree), so it finds nothing, waits 5 seconds and throws. These are the two most common ways to hide something, and both are judged fail.
+- Suggested fix: change this step to `getByRole("button", { name, includeHidden: true })`, or treat "not found" as "not shown"; the later lookups after hovering need no change.
+- Effect on existing runs: none. All 13 runs use opacity + pointer-events (measured: every diff was read).
 
 ## menu-clip
 
-要求：列表最后几行的 ⋯ 菜单完整露出，Move to Done、Delete 能点；上面的行菜单仍在 ⋯ 旁边打开。全部查到。
+Requirements: the ⋯ menu of the last rows of the list shows in full, and Move to Done and Delete can be clicked; the menus of rows higher up still open next to ⋯. All checked.
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
-| portal-no-flip（挂 body 但总往下开） | fail | OK（Delete 出了窗口） |
-| open-upward-always（就地一律往上开） | fail | OK（第一行被裁） |
-| unclip-inbox-only（只去掉外框 overflow: hidden） | fail | OK |
-| inplace-flip（不挂 body，量滚动区剩余高度，放不下就往上开） | pass | OK |
+| portal-no-flip (attached to body but always opens downwards) | fail | OK (Delete is outside the window) |
+| open-upward-always (in place, always opens upwards) | fail | OK (clipped on the first row) |
+| unclip-inbox-only (only the outer frame's overflow: hidden removed) | fail | OK |
+| inplace-flip (not attached to body; measures the height left in the scroll area and opens upwards when it does not fit) | pass | OK |
 
-判分不要求 portal，换写法不受冤枉。10 局都是 portal + 翻转。
+The judge does not require a portal, so another approach is not failed wrongly. All 10 runs use a portal + flip.
 
 ## narrow-nav
 
-要求：390px 宽无横向滚动；导航藏到页头的 Menu 按钮后面；点开盖在页面上；选页面跳过去并收起；桌面窗口侧栏（和页头）保持原样。
+Requirements: no horizontal scrolling at 390px wide; the navigation is tucked behind a Menu button in the header; opened, it lies over the page; choosing a page goes there and closes it; in a desktop window the sidebar (and the header) stay as they were.
 
-没被检查：**宽屏下页头多出 Menu 按钮**（判分在 1200px 只看侧栏链接还在左边）；Menu 按钮在不在页头里（只找页面上任意可见的、名字带 menu/navigation 的按钮）。
+Not checked: **an extra Menu button in the header on wide screens** (at 1200px the judge only looks at whether the sidebar links are still on the left); whether the Menu button is in the header (it looks for any visible button on the page whose name has menu/navigation in it).
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
-| **menu-button-on-desktop**（抽屉做对了，Menu 按钮宽屏也显示） | fail | **LEAK** |
-| inflow-drawer（导航插在页面顶部、把内容往下推） | fail | OK（判分先把任务滚进视野，顶部展开的导航不在屏上） |
-| inflow-no-close（同上且选了页面不收） | fail | OK（同上，第一步就失败） |
-| overlay-no-close（参考抽屉，选了页面不收） | fail | OK |
-| right-drawer-named（按钮叫 Open navigation，右边滑出） | pass | OK |
-| **toggle-sidebar-label**（参考答案，按钮无障碍名叫 Toggle sidebar） | pass | **STRICT**（边缘）：there is no Menu button |
+| **menu-button-on-desktop** (the drawer is right, the Menu button also shows on wide screens) | fail | **LEAK** |
+| inflow-drawer (the navigation is inserted at the top of the page and pushes the content down) | fail | OK (the judge scrolls a task into view first, and the navigation opened at the top is then off screen) |
+| inflow-no-close (same, and choosing a page does not close it) | fail | OK (same; fails at the first step) |
+| overlay-no-close (the reference drawer, not closed when a page is chosen) | fail | OK |
+| right-drawer-named (the button is called Open navigation and it slides in from the right) | pass | OK |
+| **toggle-sidebar-label** (the reference solution with the button's accessible name Toggle sidebar) | pass | **STRICT** (borderline): there is no Menu button |
 
-- LEAK menu-button-on-desktop：建议在 1200px 那段加一条：没有可见的、名字匹配 `/menu|navigation/i` 的按钮（或页头里按钮数与原来一致）。影响已有局：无（已量：10 局在 1200px 下可见按钮都只有 Notifications、Dark mode 和页面自己的按钮）。
-- STRICT toggle-sidebar-label（边缘）：题面原话是「a Menu button」，叫 Menu 的要求说得过去；若要放宽，可改成认页头里 `aria-controls` 指向导航、或 `aria-expanded` 的按钮。影响已有局：无。
+- LEAK menu-button-on-desktop: suggested fix, add to the 1200px part: no visible button whose name matches `/menu|navigation/i` (or the number of buttons in the header is the same as before). Effect on existing runs: none (measured: at 1200px the visible buttons of the 10 runs are only Notifications, Dark mode and the page's own buttons).
+- STRICT toggle-sidebar-label (borderline): the ticket's words are "a Menu button", so requiring it to be called Menu is defensible; to loosen it, accept a button in the header whose `aria-controls` points at the navigation, or one with `aria-expanded`. Effect on existing runs: none.
 
 ## dark-mode
 
-要求：1) 没选过时跟随系统；2) 选了以后刷新记得；3) 下次来访（come back later）记得；4) 收件箱深色、文字可读。
+Requirements: 1) follows the system when nothing was chosen; 2) a choice is remembered across a reload; 3) remembered on a later visit (come back later); 4) the Inbox is dark and the text readable.
 
-没被检查：
-- 「下次来访」：判分的两次「记得」都在同一个标签页里（`reload` 和 `goto`），sessionStorage 也能过。
-- 「可读」只量了主文字（标题、描述、评论正文、小节标题），次要文字（行里的优先级/Done、字段名、评论作者和时间、附件大小、任务数）没量。
+Not checked:
+- "A later visit": both of the judge's "remembered" checks happen in the same tab (`reload` and `goto`), which sessionStorage passes too.
+- "Readable" was only measured on the main text (titles, descriptions, comment bodies, section headings), not on secondary text (priority / Done in a row, field names, comment author and time, attachment size, the task count).
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
-| **session-storage**（选择记在 sessionStorage） | fail | **LEAK** |
-| no-system-default（不跟系统） | fail | OK |
-| round1-only（第 2 轮没做） | fail | OK |
-| **dim-secondary**（收件箱深色，次要文字 #39404d，约 1.6:1） | fail | **LEAK** |
-| select-picker（开关换成 System / Light / Dark 下拉） | pass | OK |
+| **session-storage** (the choice kept in sessionStorage) | fail | **LEAK** |
+| no-system-default (does not follow the system) | fail | OK |
+| round1-only (round 2 not done) | fail | OK |
+| **dim-secondary** (the Inbox is dark, secondary text #39404d, about 1.6:1) | fail | **LEAK** |
+| select-picker (the switch replaced by a System / Light / Dark dropdown) | pass | OK |
 
-- LEAK session-storage：建议选完之后在同一 context 里 `newPage()` 打开应用（新标签页共享 localStorage、不共享 sessionStorage）再看主题；或者新开 context 前把 storageState 带过去。影响已有局：无（6 局都用 localStorage，diff 看过）。
-- LEAK dim-secondary：建议在 `checks` 里加几条次要文字（列表行的 meta、字段名 dt、评论作者和时间、附件大小），门槛可放到 3:1 或 4.5:1。参考答案量出来最低是评论时间 5.2:1，4.5 也能过。影响已有局：无（已量：6 局这些文字最低 6.2:1）。
+- LEAK session-storage: suggested fix, after choosing, open the app with `newPage()` in the same context (a new tab shares localStorage and not sessionStorage) and look at the theme; or carry the storageState into a new context. Effect on existing runs: none (all 6 runs use localStorage; the diffs were read).
+- LEAK dim-secondary: suggested fix, add a few secondary texts to `checks` (the meta of a list row, the dt field names, comment author and time, attachment size), with a threshold of 3:1 or 4.5:1. The lowest measured on the reference solution is the comment time at 5.2:1, so 4.5 passes too. Effect on existing runs: none (measured: the lowest of these texts over the 6 runs is 6.2:1).
 
 ## scroll-tail
 
-要求：滚到底最后一条完整露出；新发的评论不被藏。全部查到。
+Requirements: scrolled to the end, the last item shows in full; a newly posted comment is not hidden. All checked.
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
-| calc-short（高度扣得不够） | fail | OK |
-| box-overflow-visible（去掉外框裁剪，列表压到评论框上） | fail | OK |
-| scroll-on-post（只在发评论后滚到底） | fail | OK |
-| grid-box（外框 grid，列表 minmax(0, 1fr)） | pass | OK |
+| calc-short (too little height subtracted) | fail | OK |
+| box-overflow-visible (the outer frame's clipping removed; the list runs over the comment box) | fail | OK |
+| scroll-on-post (only scrolls to the end after a comment is posted) | fail | OK |
+| grid-box (outer frame grid, list minmax(0, 1fr)) | pass | OK |
 
 ## mention-trim
 
-要求：每行只剩头像和名字；去掉邮箱；去掉 ID；名字占满一行、单行显示长名字；键盘 ↑↓ 移动、Enter/Tab 选、Esc 关。
+Requirements: each row keeps only the avatar and the name; the email is removed; the ID is removed; the name takes the whole row and a long name shows on one line; on the keyboard ↑↓ move, Enter/Tab pick, Esc closes.
 
-没被检查：**头像还在**；**↑ 键**（只按了 ↓）。
+Not checked: **the avatar still being there**; **the ↑ key** (only ↓ was pressed).
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
-| **no-avatar**（头像也删了） | fail | **LEAK** |
-| **arrowup-broken**（↑ 不再移动） | fail | **LEAK** |
+| **no-avatar** (the avatar removed too) | fail | **LEAK** |
+| **arrowup-broken** (↑ no longer moves) | fail | **LEAK** |
 | esc-broken | fail | OK |
-| no-roles-esc-broken（去掉角色且 Esc 坏） | fail | OK（但见下面 STRICT，是被别的步骤判掉的） |
-| hidden-flex（JSX 不动，邮箱/ID display: none，行改 flex） | pass | OK |
-| **no-roles**（参考答案，只去掉 listbox/option 角色） | pass | **STRICT**：typing @maxi does not show the matching member |
+| no-roles-esc-broken (roles removed and Esc broken) | fail | OK (but see the STRICT below: it is failed by another step) |
+| hidden-flex (JSX untouched, email/ID display: none, the row changed to flex) | pass | OK |
+| **no-roles** (the reference solution with only the listbox/option roles removed) | pass | **STRICT**: typing @maxi does not show the matching member |
 
-- LEAK no-avatar：建议每个 option 里要有头像（例如 option 内名字左边有一个 ≥16px 的非文字元素，或含该成员缩写的元素）。影响已有局：无（8 局 diff 都只删了邮箱和 ID 两个 span）。
-- LEAK arrowup-broken：建议键盘那段改成 ↓↓↑ 再 Enter，要求选中第二项。影响已有局：无（8 局都没动按键处理）。
-- STRICT no-roles（边缘）：`mentionPopup` 没有 listbox 时退回「同时含 Priya Shah 和 Diego Alvarez 的容器」，输入 @maxi 过滤后只剩一人，必然找不到。另外 Esc 那步只数 `role=option`，没有角色时「没关」也看不出。建议：第一次找到弹窗后记下它（或它的 id/类名），后面用同一个元素判断是否还显示。影响已有局：无（8 局都保留了角色）。
+- LEAK no-avatar: suggested fix, each option must have an avatar (for example a non-text element of ≥16px to the left of the name inside the option, or an element with the member's initials). Effect on existing runs: none (the diffs of the 8 runs only remove the two spans for the email and the ID).
+- LEAK arrowup-broken: suggested fix, change the keyboard part to ↓↓↑ and then Enter, requiring the second option to be picked. Effect on existing runs: none (none of the 8 runs touches the key handling).
+- STRICT no-roles (borderline): without a listbox, `mentionPopup` falls back to "a container that holds both Priya Shah and Diego Alvarez"; after typing @maxi the filter leaves one person and it can never be found. Also, the Esc step only counts `role=option`, so without roles "not closed" does not show either. Suggested fix: remember the popup the first time it is found (or its id / class name) and use the same element afterwards to tell whether it is still shown. Effect on existing runs: none (all 8 runs keep the roles).
 
 ## type-scale
 
-要求（约 25 条）：三档字号分别用于题面列出的各类文字；头像缩写、看板列标题、任务页头（标题、Edit、Copy link）、列表页头与筛选、其他页面不动。
+Requirements (about 25): the three sizes are used for the kinds of text the ticket lists; avatar initials, Board column headings, the task page header (title, Edit, Copy link), the list header and filters, and other pages are left alone.
 
-判分精确量了大部分点名的元素，其余只要求「在 12/13/14 里」。没被检查或只查了范围的：
-- 不动的：**头像缩写**（带 aria-hidden，`sizesIn` 直接跳过）、**任务页头的 Edit / Copy link**（`sizesIn(pane, head)` 把页头排除了，又只量了标题）、**列表页头与筛选**（Status、搜索框、任务数都没量）。
-- 只查范围的：**Done 标记应为 12**、**Comments 小节标题应为 14**、Attach file 与「Add a comment」标签应为 13。
+The judge measures most of the named elements exactly and for the rest only requires "one of 12/13/14". Not checked, or only checked for the range:
+- Left alone: **avatar initials** (they have aria-hidden, which `sizesIn` skips outright), **Edit / Copy link in the task page header** (`sizesIn(pane, head)` excludes the header, and only the title is measured), **the list header and filters** (Status, the search box and the task count were not measured).
+- Range only: **the Done mark should be 12**, **the Comments section heading should be 14**, Attach file and the "Add a comment" label should be 13.
 
-| 变体 | 期望 | 结果 |
+| Variant | Expect | Result |
 |---|---|---|
 | **avatar-12** | fail | **LEAK** |
 | **done-13** | fail | **LEAK** |
 | **comments-heading-13** | fail | **LEAK** |
 | **title-bar-buttons-13** | fail | **LEAK** |
 | **list-filters-13** | fail | **LEAK** |
-| vars-in-global（全局 styles.css 里用 CSS 变量和别的选择器写三档） | pass | OK |
+| vars-in-global (the three sizes written in the global styles.css with CSS variables and other selectors) | pass | OK |
 
-- 建议：按题面逐条补精确值——头像缩写 11px（原值）、Edit / Copy link 15px（原值）、Status 筛选 / 搜索框 15px、任务数 13px、Done 12px、Comments 标题 14px、Attach file 和 Add a comment 13px。`sizesIn` 跳过 aria-hidden 是为了不把头像算进三档，这没问题，只是要另外单独量一次头像。
-- 影响已有局：无（已量：8 局上面这些值全部符合题面）。
-- 顺带：vars-in-global 第一版判失败，是变体自己的选择器优先级写错了（附件名被 13px 覆盖），改正后通过，不是 STRICT。
+- Suggested fix: add the exact values item by item from the ticket: avatar initials 11px (the original value), Edit / Copy link 15px (the original value), the Status filter and the search box 15px, the task count 13px, Done 12px, the Comments heading 14px, Attach file and Add a comment 13px. `sizesIn` skips aria-hidden so as not to count the avatar among the three sizes, which is fine; the avatar just has to be measured once on its own.
+- Effect on existing runs: none (measured: on the 8 runs all of these values agree with the ticket).
+- In passing: the first version of vars-in-global was judged fail because the variant's own selector specificity was wrong (the attachment name was overridden by 13px); corrected, it passes. Not a STRICT.
 
 ---
 
-## 最该先改的 5 处
+## The five things to fix first
 
-1. **hover-ghost：`getByRole` 找不到 display: none / visibility: hidden 的按钮，判分抛错**（STRICT ×2）。最常见的两种隐藏写法都被冤枉。第一步改 `includeHidden: true` 或把「找不到」当成「没显示」。
-2. **dark-mode：「下次来访还记得」没测**（LEAK）。sessionStorage 能过。选完后 `context.newPage()` 再看主题。顺手把次要文字的对比度也量上（LEAK dim-secondary）。
-3. **type-scale：「不动」的几处和 Done / Comments 标题没精确量**（LEAK ×5）。按题面逐条补精确值，单独量一次头像缩写。
-4. **mention-trim：头像和 ↑ 键没查**（LEAK ×2）。都是题面明写、判分一句话能补上的。
-5. **narrow-nav：宽屏多出 Menu 按钮没查**（LEAK）。同样一句话能补上。
+1. **hover-ghost: `getByRole` does not find a button with display: none / visibility: hidden, and the judge throws** (STRICT ×2). The two most common ways to hide something are both failed wrongly. Change the first step to `includeHidden: true`, or treat "not found" as "not shown".
+2. **dark-mode: "still remembered on a later visit" is not tested** (LEAK). sessionStorage passes. After choosing, `context.newPage()` and look at the theme. Measure the contrast of secondary text along the way (LEAK dim-secondary).
+3. **type-scale: the "leave alone" places and the Done / Comments heading are not measured exactly** (LEAK ×5). Add the exact values item by item from the ticket and measure the avatar initials once on their own.
+4. **mention-trim: the avatar and the ↑ key are not checked** (LEAK ×2). Both are stated in the ticket and take one line each in the judge.
+5. **narrow-nav: an extra Menu button on wide screens is not checked** (LEAK). Likewise one line.
 
-其次：loading-veil 首次加载的转圈（LEAK）、mention-trim 弹窗退回逻辑（STRICT 边缘）、loading-veil 用 `isVisible` 不看 opacity（STRICT 边缘）、narrow-nav 只认名字带 menu/navigation 的按钮（STRICT 边缘）。
+Next: the spinner on the first load of loading-veil (LEAK), the popup fallback of mention-trim (STRICT, borderline), `isVisible` not looking at opacity in loading-veil (STRICT, borderline), narrow-nav accepting only a button with menu/navigation in its name (STRICT, borderline).
 
-以上判分改动都只会让「漏掉题面要求」的实现判失败、或让换写法的对实现判通过；按已量的数据和 diff，**已有局的结论都不会变**。
+All of these judge changes only make implementations that "miss something the ticket asks for" fail, or make correct implementations done another way pass; by the measured data and the diffs, **the verdict of no existing run changes**.
 
 ---
 
-## 修订结果（按报告全部改）
+## What was changed (everything in the report)
 
-改了 6 题，判分代码只动 `tasks/interaction.ts`、`tasks/layout.ts`，没动共享文件、Team Tasks、参考答案、种子。
-最后一次全量 `npx tsx harness/audit.ts -j 1`（11 题）：**55/55 OK**；6 题 `harness/selftest.ts` 全部「种子判失败、参考答案判通过」。
-抽查重判用只读的重判脚本（不写回结果文件），**共 30 局，全部仍判通过，没有结论变化**。
+6 tasks were changed. The judge code is only touched in `tasks/interaction.ts` and `tasks/layout.ts`; shared files, Team Tasks, the reference solutions and the seeds were not touched.
+The last full `npx tsx harness/audit.ts -j 1` (11 tasks): **55/55 OK**; `harness/selftest.ts` for the 6 tasks all gives "the seed is judged fail, the reference solution pass".
+The spot re-judge used the read-only re-judge script (which does not write back to the result files): **30 runs in all, every one still judged pass, no verdict changed**.
 
-| 题 | 改了什么 | 变体 | 抽查重判（只读） |
+| Task | What changed | Variants | Spot re-judge (read-only) |
 |---|---|---|---|
-| loading-veil | 收紧：首次打开收件箱（慢网络）1.5s 内要有加载提示。放宽：加载提示是否显示改用 checkVisibility（看 opacity） | 5/5 OK（no-initial-spinner 判失败，fade-with-spinner 判通过） | ChatGPT Computer Use 第 1、3 局，Mirasim Computer Use 第 4、5 局通过 |
-| hover-ghost | 放宽：找 ✓/☆ 用 includeHidden；点 ☆ 前先悬停到那一行 | 7/7 OK，见下 | ChatGPT Computer Use 第 1、5 局，Mirasim Computer Use 第 2 局和一局试跑通过 |
-| narrow-nav | 收紧：1200px 时窄屏那个菜单按钮不能还显示。放宽：菜单按钮名字带 sidebar、或 aria-controls 指向主导航的也认 | 6/6 OK | ChatGPT Computer Use 第 1 局和一局试跑，Mirasim Computer Use 两局试跑通过 |
-| dark-mode | 收紧：选浅色并刷新后，新开标签页（系统仍深色）也要是浅色；深色下 6 处次要文字对比度 ≥ 3:1 | 5/5 OK | ChatGPT Computer Use 第 1–3 局、Mirasim Computer Use 第 1–3 局通过 |
-| mention-trim | 收紧：每行名字左边要有头像；键盘改成 ↓↓↑ 回车选第二项。放宽：没有 listbox 时记住第一次找到的弹窗（标签 + 类名）再找；Esc 是否关掉也看这个弹窗 | 6/6 OK | ChatGPT Computer Use 第 1、2 局，Mirasim Computer Use 第 2 局和一局试跑通过 |
-| type-scale | 收紧：Done 12、Comments 标题 14、Attach file 与 Add a comment 13 改成精确值；头像缩写 11、任务页头 Edit/Copy link 15、Status 筛选与搜索框 15、任务数 13 要求保持原值 | 6/6 OK | ChatGPT Computer Use 第 1–3 局，Mirasim Computer Use 第 1–3 局和两局试跑通过（8 局全判） |
-
-没改的 5 题：divider-lock、card-stretch、sticky-save、menu-clip、scroll-tail，审计没有发现 LEAK/STRICT。
-
-**hover-ghost 的样例变体结论变了**：判分修好、不再抛错之后，display-none-hover 和 visibility-hidden 都判
-「点任务名打不开（点在 div.row-actions 上）」。原因是这两个变体只换了隐藏方式，悬停时那个整行宽的透明容器
-照样显示、盖住任务名，工单的 bug 根本没修。所以它们是错的实现，why.md 改成 expect fail 并写明理由；审计时
-记的这 2 处 STRICT，实质是判分抛错（includeHidden 的问题确实存在，已改）。另加 display-none-compact、
-visibility-compact（容器 pointer-events: none、按钮 auto，再用 display / visibility 隐藏）作为真正对的
-换写法，旧判分会因找不到按钮抛错，新判分判通过。
+| loading-veil | Stricter: opening the Inbox for the first time (slow network) must show a loading indicator within 1.5s. Looser: whether the loading indicator is shown now uses checkVisibility (which looks at opacity) | 5/5 OK (no-initial-spinner is judged fail, fade-with-spinner pass) | ChatGPT Computer Use runs 1 and 3, Mirasim Computer Use runs 4 and 5 pass |
+| hover-ghost | Looser: ✓/☆ are found with includeHidden; the row is hovered before ☆ is clicked | 7/7 OK, see below | ChatGPT Computer Use runs 1 and 5, Mirasim Computer Use run 2 and a trial run pass |
+| narrow-nav | Stricter: at 1200px the narrow-screen menu button must not still show. Looser: a menu button with sidebar in its name, or with aria-controls pointing at the main navigation, is accepted too | 6/6 OK | ChatGPT Computer Use run 1 and a trial run, two Mirasim Computer Use trial runs pass |
+| dark-mode | Stricter: after choosing light and reloading, a new tab (with the system still dark) must be light too; in dark, the contrast of 6 pieces of secondary text is ≥ 3:1 | 5/5 OK | ChatGPT Computer Use runs 1–3, Mirasim Computer Use runs 1–3 pass |
+| mention-trim | Stricter: each row must have an avatar to the left of the name; the keyboard part is now ↓↓↑ and Enter picking the second option. Looser: without a listbox, the popup found the first time is remembered (tag + class name) and looked up again; whether Esc closed it is judged on that popup too | 6/6 OK | ChatGPT Computer Use runs 1 and 2, Mirasim Computer Use run 2 and a trial run pass |
+| type-scale | Stricter: Done 12, the Comments heading 14, and Attach file and Add a comment 13 are now exact values; avatar initials 11, Edit / Copy link in the task page header 15, the Status filter and the search box 15 and the task count 13 must keep their original values | 6/6 OK | ChatGPT Computer Use runs 1–3, Mirasim Computer Use runs 1–3 and two trial runs pass (all 8 runs judged) |
+
+The 5 tasks not changed: divider-lock, card-stretch, sticky-save, menu-clip, scroll-tail; the audit found no LEAK / STRICT in them.
+
+**The verdict on hover-ghost's sample variants changed**: once the judge was fixed and no longer throws, display-none-hover and visibility-hidden are both judged "clicking the task's name does not open it (the click lands on div.row-actions)". The cause is that these two variants only change how things are hidden; on hover the transparent container as wide as the row still shows and covers the task's name, so the ticket's bug is not fixed at all. They are therefore wrong implementations, and their why.md was changed to expect fail with the reason given; the 2 STRICTs recorded in the audit were in substance the judge throwing (the includeHidden problem is real and is fixed). display-none-compact and visibility-compact were added (container pointer-events: none, buttons auto, then hidden with display / visibility) as the really correct alternatives: the old judge throws because it cannot find the button, and the new judge passes them.

docs/judging.md

@@ -28,7 +28,7 @@ Each task has a script in `judges/real-projects/` that takes the dev server's UR
 | Real projects, fix | `... --dry --fixed` | The upstream fix (`git diff base merge -- fixPaths`, source files only) must pass |
 | Real projects, live fix | `... --dry --fixed --live` | The fix applied while the dev server runs must pass after hot reload, so a fix made by the agent is visible to the judge |
 
-The audit was done task by task: list every requirement in the ticket, check which ones the judge tests, write variants for the gaps, fix the judge, and re-read the real runs for verdicts that would change. Notes per group of tasks (in Chinese): [`docs/judge-audits/`](judge-audits/).
+The audit was done task by task: list every requirement in the ticket, check which ones the judge tests, write variants for the gaps, fix the judge, and re-read the real runs for verdicts that would change. Notes per group of tasks: [`docs/judge-audits/`](judge-audits/).
 
 ## Re-judging