fix(rectification): match the delivery inside the acknowledged reply (BUG-1241)

The answered choice persists the acknowledgement, a blank line, and the
rewritten delivery. The exit fill-in required equality and wrote the
deterministic delivery again. Check containment instead, and pin it on the
real answer path. The suspected widen staleness regression did not hold;
keep a guard test.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01N4f2nya58RoRu4yEmJgRGE

docs/BUG_HISTORY.md

@@ -16677,17 +16677,17 @@
 ## BUG-1241 | 校正交付后同一轮又补出第二段回答
 
 - 状态：resolved（分支 `codex/rectification-delivery-dup-adopt-20261006`）
-- 首次发现 / 最近更新：2026-10-06 / 2026-10-06
-- 来源：10-06 测试环境真机。任务书 `docs/tasks/TASK-rectification-delivery-dup-adopt-20261006.md`。
+- 首次发现 / 最近更新：2026-10-06 / 2026-10-06（验收修复单后）
+- 来源：10-06 测试环境真机。任务书 `docs/tasks/TASK-rectification-delivery-dup-adopt-20261006.md`；验收修复单 `docs/tasks/TASK-rectification-delivery-dup-adopt-fix-20261006.md`。
 - 影响面：生时校正答完最后一题后的交付。不改引擎，不改数据库。
 - 现象：同一轮出现两段助手回答。前一段是采用旁白的改写。后一段是确定性兜底，带分钟范围和「还剩几个候选」。刷新后两段都在。
 - 触发条件：答题路径已经把改写后的交付句写入本轮；收尾补位找不到已问轮次编号，又用子串去比对。改写句不是兜底原文的子串，于是再写一条。
 - 根因：BUG-1153 的防重只看「上一条是否包含确定性原文」。当时的测试把上一条写成已经包含这段原文，所以没覆盖「上一条是另一句改写」。60 秒内不重复记账的标记也没有拿来比对这句已经写下的交付。
-- 修复：答题回复在落库前记下这句交付原文。收尾补位若发现上一条助手原文（去掉空白后）就是这句，就不再写。子串比对保留。没有任何回复承载这句时，补位仍写 1 条。不把旁白函数传进收尾，也不改旁白提示词。选择题 JSON 在已有轮次编号时带上 `x-rectification-turn-id`，作为额外保险；真库验收不依赖这个头。
-- 验证：`frontend/tests/rectification-delivery-dup-adopt-20261006.test.ts` 覆盖「上一条是改写句则 0 条新写入」和「上一条不是这句则仍写 1 条」。BUG-1153 原测试仍通过。真 PostgreSQL 17 上，种好的并列卷宗走同一条收尾后，助手交付消息恰好 1 条。本机没有星历引擎，收尾里的刷新失败得很快，失败之后也没有写出第 2 条。
-- 防复发：上述测试。跳过条件是「上一条等于已记下的交付句」，不能只靠加长子串。
+- 修复：答题回复在落库前记下这句交付原文。收尾补位若发现上一条助手原文（去掉空白后）就是这句，就不再写。子串比对保留。没有任何回复承载这句时，补位仍写 1 条。**验收修复（b977f482 之后）**：首版要求上一条「等于」这句，但答题回复真实落库是「已记录，…」+ 空行 + 改写句（`persistApplied` 在 `skippedNextInterview` 时拼接答题回执），于是事故原样仍补出第二条；改为「上一条包含这句」（去空白比子串，不足 8 字不匹配）。不把旁白函数传进收尾，也不改旁白提示词。选择题 JSON 在已有轮次编号时带上 `x-rectification-turn-id`，作为额外保险；真库验收不依赖这个头。
+- 验证：`frontend/tests/rectification-delivery-dup-adopt-20261006.test.ts` 覆盖「上一条是改写句则 0 条新写入」和「上一条不是这句则仍写 1 条」；修复单补「回执前缀 + 改写」两种回执各 0 条、「只有回执」仍 1 条。`frontend/tests/rectification-answer-choice.test.ts`「a rewritten delivery reply is not followed by a second exit fill-in turn」走真实答题路径（`applyRectificationChoice` + 改写旁白）取实际落库文本再调收尾：b977f482 上 1 条（红），修后 0 条。BUG-1153 原测试仍通过。真 PostgreSQL 17 上，种好的并列卷宗走同一条收尾后，助手交付消息恰好 1 条。本机没有星历引擎，收尾里的刷新失败得很快，失败之后也没有写出第 2 条。
+- 防复发：上述测试。跳过条件是「上一条包含已记下的交付句」，比对对象必须取自真实答题路径的落库文本，不得手写只含改写句的夹具（首版回放就是这样漏掉回执前缀的）。
 - 相关记录：BUG-596、BUG-1149、BUG-1153。
-- 复发自：BUG-1153。旧测试的上一条已经包含确定性原文，改写句走不到那条断言。
+- 复发自：BUG-1153。旧测试的上一条已经包含确定性原文，改写句走不到那条断言。首版修复的测试与真库回放都只写了改写句、没有答题回执前缀。
 - 修复版本：待发布
 
 ## BUG-1242 | 点采用被拒，因为采用和读取没用答题时的出生日期

docs/tasks/PROGRESS-rectification-delivery-dup-adopt-20261006.md

@@ -105,3 +105,48 @@ T2 做法：不把采用旁白函数传进收尾，也不改旁白提示词。
 `next build --webpack`。Turbopack 会拒绝工作树里指到主检出的 `node_modules` 联接，所以用 webpack。基线 `65b61ed5` 编译 2.9 分钟、构建自带的 TypeScript 100 秒，都通过，收集页面数据时停在 `/api/rectification/agent`。本分支编译 2.2 分钟、TypeScript 83 秒，都通过，收集页面数据时停在 `/api/birth-time-guide`。两边都是 Windows 不允许为 Skill 运行别名建符号链接（`EPERM`）。先碰到哪条路由取决于收集页面的工人，不是这次改动。路由表没有打出来，这台机器没能确认 `/` 仍是 Static，也没有首屏 gzip 数字。没有为了构建改符号链接，也没有在工作树里重装依赖。首页 `page.tsx` 没改；校正卡片组件有改，gzip 要等能建符号链接的环境再量。
 
 本轮没有打开浏览器，也没有在手机上点采用。真机步骤在 `docs/testing/rectification-delivery-dup-adopt-20261006.md`。修复版本待发布，没有推送，没有部署。
+
+## 验收修复单（Claude 直接执行，2026-10-06）
+
+修复单：`docs/tasks/TASK-rectification-delivery-dup-adopt-fix-20261006.md`。在本分支 `b977f482` 之上追加。Linux，Node 22.14。
+
+### F1 防重改为包含判断（BUG-1241 重开后修复）
+
+- 事实：答题回复真实落库是「答题回执 + 空行 + 改写句」。`persistApplied` 在交付轮 `skippedNextInterview=true` 时拼 `${input.narration}\n\n${hostNarration}`。首版 `deliveryCarrierMatchesLastAssistant` 要求全等，事故原样仍补第二条。
+- 修：`delivery-turn-guard.ts` 改为 `text.includes(needle)`（去空白，不足 8 字不匹配）。
+- 新测试（均在 `b977f482` 上红、修后绿）：
+  - `rectification-delivery-dup-adopt-20261006.test.ts`：「已记录，范围没变。」+ 改写 → 0 条；「已记录，范围没变；…领先…落后。」+ 改写 → 0 条；只有回执 → 仍 1 条。
+  - `rectification-answer-choice.test.ts`「a rewritten delivery reply is not followed by a second exit fill-in turn」：走 `applyRectificationChoice` 真实答题路径并注入改写旁白，取 `append_agentic_rectification_turn` 实际写入的文本（实测为「已记录，范围收到 …」+ 空行 + 改写），再调 `persistExhaustionGateTurn`：`b977f482` 上写入 1 条，修后 0 条。
+- 真库：本轮**没有**做 PG17 重放（修复单要求）。替代证据是上面这条：落库文本取自生产拼接代码，而不是手写夹具，首版恰恰是手写夹具漏掉了回执前缀。改动只是内存比较，不涉及 SQL。PG17 重放记为欠项。
+
+### F2 撤回（修复单 §2.2 判断有误，BUG-1246 不成立，不入 Bug 历史）
+
+修复单怀疑 `persistApplied` 删掉 `snapshotCurrent: input.snapshotCurrent` 后，放宽窗口失败时会按旧窗口出卡。复核结果：
+
+- 放宽窗口（`mutateCaseForWidenWindow`）与选时段（`mutateCaseForBlockChoice`）进 `persistApplied` 时 `decisionState: null`。`decideAfterInferenceChange` 在 state 为空时 `candidateScores: []`，一律 `collect_evidence`，snapshotCurrent 不参与。
+- 唯一带 state 的调用（stop，`snapshotCurrent: rescored.snapshotCurrent`）里，这个值本来就是 `scoreableSnapshotCurrentFromDossier` 对同一卷宗算出来的。stored 与 current 只差证据指纹，revision 两边抵消，与函数内自算结果相同。
+- 结论：删那一行不改变行为。保留一条守护测试「answering widen A with a failed rescore never delivers or adopts on the old window's candidates」（对照组同一份存量快照按当前读是可采用的并列交付；放宽后 `can_adopt=false`、不出 `ready_to_adopt` / `complete_with_range`），修前修后都绿。BUG-1246 号码不使用。
+
+### 本机测试（对照基线 `86cafe68`，同机同命令）
+
+| | tests | pass | fail | cancelled |
+| --- | --- | --- | --- | --- |
+| 基线 `86cafe68` | 4936 | 4871 | 24 | 0 |
+| `b977f482` | 4952 | 4886 | 25（多 1 条首页计时 flaky，单跑 17/17） | 0 |
+| 本轮 | 4957 | 4892 | 24 | 0 |
+
+失败名与基线逐名相同（diff 为空）。新增 21 条，没有删除。`tsc --noEmit` 0；`npm run lint` 0 error，126 warning（与基线相同）。
+`next build`：本轮只改了服务端的 `delivery-turn-guard.ts` 和测试，没有重建。`b977f482` 的构建 Claude 已验：`/` 是 `○ Static`，首屏 gzip 636,730 对 636,565（+0.03%）。
+
+既有断言：未改。
+
+### 变基到 staging `da226db1` 后复验
+
+staging 在这期间进了普通对话的代码（`da226db1`），本分支已变基到它之上。CHANGELOG、BUG_HISTORY、README 的冲突都是两边各追加，按「BUG 号升序」「两条 CHANGELOG 都留」解决。复验以 `da226db1` 为基线，同机同命令：
+
+| | tests | fail | cancelled |
+| --- | --- | --- | --- |
+| 基线 `da226db1` | 4941 | 24 | 0 |
+| 本分支（变基后） | 4962 | 24 | 0 |
+
+失败名逐名相同，测试名无消失，新增 21 条。tsc 0；lint 0 error。`next build`：`/` 是 `○ Static`；首屏 gzip 636,129 对 635,958（+171 B，+0.03%）。`tests/test_repo_privacy_markers.py` 通过。

docs/tasks/README.md

@@ -423,5 +423,5 @@
 | `TASK-gate-log-volume-20261004.md` | `PROGRESS-gate-log-volume-20261004.md` | 门禁 validate 单步日志 4.4 万行 / 2 MB 网页打不开（run 3170）：快速门只打摘要（失败给末 200 行 + 日志文件）、前端测试门禁上 dot + 失败汇总（本机仍 TAP）、拆分 validate 为 7 个 step（产品授权改 workflow，只限拆分与重定向）；检查一项不少 | **已实现待验收**（BUG-1230；未推送、未部署；Gitea 各 step 页面是否打得开留待推送后由产品确认） | 分支 `codex/gate-log-volume-20261004` |
 | `TASK-consult-latency-quickwins-20261005.md` | `PROGRESS-consult-latency-quickwins-20261005.md` | 普通对话耗时两项快修：咨询链校正闸不再同步白等 VedAstro 官方快照子进程（每域约 4 s，复用顶层缓存 + 负缓存，不推翻 BUG-301）；补分段计时（第 0/1 步耗时、推理 token、分类耗时进日志与 usage）。推理强度/精简说明待模型 key 另单（BUG-1231、BUG-1232） | 已验收（Claude 10-05 Linux：Python 62 / 前端 24 失败与基线逐名相同，Static、gzip 0%；装 SDK 实测每领域 4.47→1.66 s，剩余 1.33 s 是保留的顶层前台等待；/admin/usage 真机欠） | 分支 `codex/consult-latency-quickwins-20261005` |
 | `TASK-rectification-nadi-seconds-research-20261005.md` | `PROGRESS-rectification-nadi-seconds-research-20261005.md` | **「纳迪秒级校准」可证伪检验（离线）**：竞品宣传「问前事到天 → 秒级」。本仓主链只用三层小运，Sookshma / Prana 与 D150 从未进评价集；v5 真值 52/77 是整 5 分钟（秒级无真值可对）。N0 五层小运 + D150 底座（前三层与主链对账 0 差）、N1 拟合率 vs 安慰剂日期（核心）、N2 留一件预测、N3 六题后区间内再细分能否提头名、N4 岁差 / 坐标 / 时间扰动的噪声地板、N5 D150 结构层（原文比对 blocked）、N6 结论 + 对外口径草稿。规则先登记再跑；不改生产代码；不得重调 BUG-1091 已关的权重 | 已验收（Claude 10-05：N1/N2/N3/N5 不过门、N4 触发文案禁令；Linux 复跑内容 0 差，结果改 LF；补 Lahiri/KP/True Chitra 对照，月亮差 0.83′ 第 5 层仍换 94%）→ 不立实现单 | 分支 `codex/rectification-nadi-seconds-research-20261005`（BUG-1240 `closed_by_design`） |
-| `TASK-rectification-delivery-dup-adopt-20261006.md` | `PROGRESS-rectification-delivery-dup-adopt-20261006.md` | **校正交付一次两段回答 + 点「改用…盘解读」被拒 `adoption_not_allowed`**（10-06 真机）：收尾补位按文字子串防重，遇采用旁白 Agent 改写版失效（BUG-1153 复发）；accept/GET 调 `decideFromDossier` 不带出生日期，未成年探针复活致判定与出卡时相反（BUG-598/680 同族）；卡片按钮不看公开 can_adopt、拒绝码直出英文。不放宽采用门，判定入参统一 + 结构防重 + 按钮判据 + 中文提示（BUG-1241～1243） | **验收未通过 → 修复单待领取**（Claude 10-06 验收 `b977f482`：T1 出生日期统一、T3 按钮/中文通过；T2 防重只认全等，事故原样「已记录，范围没变。」+改写仍补第二条；`persistApplied` 丢了放宽窗口/选时段后强制的 snapshotCurrent=false（BUG-1246）。tsc 0、lint 0 error、npm 失败与基线同名（+1 首页计时 flaky）、Static、gzip +0.03%） | 分支 `codex/rectification-delivery-dup-adopt-20261006`；修复单 `TASK-rectification-delivery-dup-adopt-fix-20261006.md` |
+| `TASK-rectification-delivery-dup-adopt-20261006.md` | `PROGRESS-rectification-delivery-dup-adopt-20261006.md` | **校正交付一次两段回答 + 点「改用…盘解读」被拒 `adoption_not_allowed`**（10-06 真机）：收尾补位按文字子串防重，遇采用旁白 Agent 改写版失效（BUG-1153 复发）；accept/GET 调 `decideFromDossier` 不带出生日期，未成年探针复活致判定与出卡时相反（BUG-598/680 同族）；卡片按钮不看公开 can_adopt、拒绝码直出英文。不放宽采用门，判定入参统一 + 结构防重 + 按钮判据 + 中文提示（BUG-1241～1243） | **已验收，待推 staging**（Claude 10-06：首版 `b977f482` 验收 T1/T3 过、T2 未过；修复单 F1 由 Claude 直接执行——防重改为「上一条包含交付句」，真实答题路径测试修前红修后绿；F2 疑似 BUG-1246 复核不成立已撤回、留守护测试。全量 npm 失败名与基线逐名相同、tsc 0、lint 0 error；PG17 重放未做，见 PROGRESS） | 分支 `codex/rectification-delivery-dup-adopt-20261006`；修复单 `TASK-rectification-delivery-dup-adopt-fix-20261006.md` |
 | `TASK-consult-conversational-answer-20261006.md` | `PROGRESS-consult-conversational-answer-20261006.md` | **普通对话首轮去汇报骨架，改成聊天**（10-06 产品反馈「像机器在汇报」）：首轮取消全部 `##`（三个固定节 + 按对象标题，多人改分段落点名）、行动不强制（盘上有具体指向才顺口一句，问「怎么办」再展开）、首轮三到六段；示范删可抄的行动句（实测被逐字照抄）；思考栏去「这周可以做什么」。推翻 10-01 D2 与 D8 首轮部分（BUG-1244～1245） | 已验收 | 28113fa0 验收未过 → 修复单 `TASK-consult-conversational-answer-fix-20261006.md` 由 Claude 直接执行（含去掉助手头像），复验通过后推 staging；真机清单 10 步待产品 |

frontend/src/lib/rectification-agentic/v9/delivery-turn-guard.ts

@@ -71,7 +71,9 @@ export function deliveryCarrierNarration(input: {
 }
 
 /**
- * True when the latest assistant text is the stored delivery sentence.
+ * True when the latest assistant text carries the stored delivery sentence.
+ * The answered choice persists 「已记录，…」 + a blank line + that sentence,
+ * so this is a containment check, not equality (BUG-1241 fix sheet).
  * Whitespace is ignored. A needle shorter than 8 characters does not match.
  */
 export function deliveryCarrierMatchesLastAssistant(input: {
@@ -91,7 +93,7 @@ export function deliveryCarrierMatchesLastAssistant(input: {
     if (turn?.role !== "assistant") continue;
     const text = typeof turn.text === "string" ? squash(turn.text) : "";
     if (!text) continue;
-    return text === needle;
+    return text.includes(needle);
   }
   return false;
 }

frontend/tests/rectification-answer-choice.test.ts

@@ -2,7 +2,8 @@ import assert from "node:assert/strict";
 import { readFileSync } from "node:fs";
 import test from "node:test";
 
-import { applyRectificationChoice } from "../src/lib/rectification-agentic/v9/answer-choice.ts";
+import { applyRectificationChoice, persistExhaustionGateTurn } from "../src/lib/rectification-agentic/v9/answer-choice.ts";
+import { resetDeliveryTurnGuardForTests } from "../src/lib/rectification-agentic/v9/delivery-turn-guard.ts";
 import { composeCollectSpokenAssistantText } from "../src/lib/rectification-agentic/v9/collect-prompt.ts";
 import { mergeTurnQuestions, type SnapshotTurnMessage } from "../src/lib/rectification-snapshot-messages.ts";
 import { choiceCardFromCaseDossier } from "../src/lib/rectification-agentic/v9/interview-state.ts";
@@ -42,6 +43,7 @@ import {
   CASE_ID,
   EVIDENCE_ID,
   FOCUS_ID,
+  RESULT_ID,
   SESSION_ID,
   TURN_ID,
   USER_ID,
@@ -734,6 +736,53 @@ test("last structured choice emits the adoption range and persists the same narr
   assert.doesNotMatch(String(turn?.args.p_assistant_message ?? ""), /排盘用/);
 });
 
+// BUG-1241 fix sheet: drive the real answer path with the adopt-narration
+// rewrite, take the assistant text it actually persists, and make sure the
+// exit fill-in does not add the deterministic delivery underneath it.
+test("a rewritten delivery reply is not followed by a second exit fill-in turn", async () => {
+  resetDeliveryTurnGuardForTests();
+  const rewrite = "因为再问下去也分不开这两个时间，这一轮到这里停住，下面是这次的结果。";
+  const accounting = persistChoiceAccounting(adoptionDossier());
+  const applied = await applyRectificationChoice(accounting.client, {
+    userId: USER_ID,
+    caseId: CASE_ID,
+    sessionId: SESSION_ID,
+    actionId: ACTION_ID,
+    action: CHOICE_ACTION,
+    focusId: FOCUS_ID,
+    questionId: QUESTION_ID,
+    probeId: RELOCATION_2015_PROBE.id,
+    optionId: "A",
+    expectedRevision: adoptionInference().revision,
+    narrateAdopt: async () => rewrite,
+  });
+  assert.equal(applied.nextAction.can_adopt, true);
+  const answerTurn = accounting.calls.find((call) => call.fn === "append_agentic_rectification_turn");
+  const persisted = String(answerTurn?.args.p_assistant_message ?? "");
+  assert.match(persisted, new RegExp(rewrite));
+
+  const deterministic = String(applied.narration).includes(rewrite) ? "" : String(applied.narration);
+  const fillIn = deterministic || "目前范围 04:45–05:15。这只是代表性候选，不是已确认的唯一出生分钟。";
+  const gate = fakeAccounting({
+    get_agentic_rectification_case_dossier: () => dossierFixture({
+      turns: [
+        { id: "t-user", role: "user", text: "A", status: "completed", created_at: "2026-10-06T00:00:00.000Z" },
+        { id: "t-answer", role: "assistant", text: persisted, status: "completed", created_at: "2026-10-06T00:00:01.000Z" },
+      ],
+    }),
+    append_agentic_rectification_turn: () => ({ turn_id: "44444444-4444-4444-8444-444444444444", idempotent: false }),
+  });
+  await persistExhaustionGateTurn({
+    accounting: gate.client,
+    userId: USER_ID,
+    caseId: CASE_ID,
+    hostNarration: fillIn,
+    resultId: RESULT_ID,
+  });
+  assert.equal(gate.calls.filter((call) => call.fn === "append_agentic_rectification_turn").length, 0);
+  resetDeliveryTurnGuardForTests();
+});
+
 test("clicking A applies the choice without invoking a language model", async () => {
   const accounting = choiceAccounting();
   const applied = await applyRectificationChoice(accounting.client, {

frontend/tests/rectification-delivery-dup-adopt-20261006.test.ts

@@ -14,7 +14,10 @@ import { RectificationRangeDelivery } from "../src/components/rectification-rang
 import { RectificationSegmentDelivery } from "../src/components/rectification-segment-delivery.tsx";
 import type { RectificationCandidateResult } from "../src/lib/rectification-candidate-result.ts";
 import type { RangeDeliveryProjection } from "../src/lib/rectification-agentic/v9/divergence-panel.ts";
-import { persistExhaustionGateTurn } from "../src/lib/rectification-agentic/v9/answer-choice.ts";
+import { applyRectificationChoice, persistExhaustionGateTurn } from "../src/lib/rectification-agentic/v9/answer-choice.ts";
+import { CHOICE_ACTION } from "../src/lib/rectification-agentic/v9/choice-action.ts";
+import { WIDEN_WINDOW_INTENT, WIDEN_WINDOW_KIND } from "../src/lib/rectification-agentic/v9/block-scan.ts";
+import { buildWidenWindowFrame } from "../src/lib/rectification-agentic/v9/window-widen.ts";
 import { dossierResponse } from "../src/lib/rectification-agentic/v9/case-dossier-response.ts";
 import { decisionInputsFromSnapshot } from "../src/lib/rectification-agentic/v9/decision-inputs.ts";
 import { decideFromDossier } from "../src/lib/rectification-agentic/v9/decision-from-dossier.ts";
@@ -32,9 +35,12 @@ import {
 } from "../src/lib/rectification-agentic/v9/tool-service.ts";
 import { adoptionRefusalCopy, runRectificationCandidateAccept } from "../src/lib/rectification-chat-accept-run.ts";
 import {
+  activeFocusFixture,
   CASE_DOSSIER_RPC,
   CASE_ID,
   candidateSnapshotFixture,
+  FOCUS_ID,
+  SESSION_ID,
   dossierFixture,
   fakeAccounting,
   RESULT_ID,
@@ -135,7 +141,12 @@ function inference(withProbe: boolean, tied: boolean) {
   };
 }
 
-function parsed(withProbe: boolean, tied: boolean, fingerprint: string | null): V9CaseDossier {
+function rawDossier(
+  withProbe: boolean,
+  tied: boolean,
+  fingerprint: string | null,
+  extra: { candidateRange?: { start_time: string; end_time: string }; activeFocus?: unknown } = {},
+) {
   const supportA = tied ? 16 : 40;
   const supportB = tied ? 16 : 8;
   const snapshot = candidateSnapshotFixture({
@@ -162,10 +173,28 @@ function parsed(withProbe: boolean, tied: boolean, fingerprint: string | null):
   const rpc = dossierFixture({
     evidenceCount: evidence.length,
     evidence,
-    candidateRange: { start_time: OPEN_START, end_time: OPEN_END },
+    candidateRange: extra.candidateRange ?? { start_time: OPEN_START, end_time: OPEN_END },
     latestResult: snapshot,
+    ...(extra.activeFocus ? {
+      conversationSummary: {
+        confirmed_evidence_summary: [],
+        pending_revisions: [],
+        active_focus: extra.activeFocus,
+        declined_skipped_topics: [],
+        candidate_divergence_summary: null,
+        missing_evidence_categories: [],
+        last_result_policy: null,
+        summary_version: 1,
+        updated_at: "2026-10-06T10:00:00.000Z",
+      },
+    } : {}),
   });
   (rpc.case as { rectification_domain?: string }).rectification_domain = "general";
+  return rpc;
+}
+
+function parsed(withProbe: boolean, tied: boolean, fingerprint: string | null): V9CaseDossier {
+  const rpc = rawDossier(withProbe, tied, fingerprint);
   const dossier = parseV9CaseDossier(rpc);
   if (!dossier) throw new Error("dossier did not parse");
   return dossier;
@@ -547,3 +576,161 @@ test("a refused adoption refreshes the snapshot and says so in Chinese", async (
   assert.equal(adoptionRefusalCopy({ error: "adoption_not_allowed" }), "这次的结果还不能采用，已刷新到最新状态");
   assert.doesNotMatch(adoptionRefusalCopy({ code: "adoption_not_allowed" }), /adoption_not_allowed/);
 });
+
+// ---------------------------------------------------------------------------
+// Fix sheet 2026-10-06 (F1, BUG-1241 reopened): the answered choice persists
+// 「已记录，…」 + blank line + the rewrite, never the rewrite alone.
+// ---------------------------------------------------------------------------
+
+for (const ack of ["已记录，范围没变。", "已记录，范围没变；04:53 仍领先，05:00 落后。"]) {
+  test(`the exit fill-in does not write again when the reply is 「${ack}」 + the stored rewrite`, async () => {
+    resetDeliveryTurnGuardForTests();
+    markDeliveryTurn({ caseId: CASE_ID, resultId: RESULT_ID, narration: REWRITE });
+    const accounting = gateAccounting(`${ack}\n\n${REWRITE}`);
+    await persistExhaustionGateTurn({
+      accounting: accounting.client,
+      userId: USER_ID,
+      caseId: CASE_ID,
+      hostNarration: FALLBACK,
+      resultId: RESULT_ID,
+    });
+    assert.equal(accounting.calls.filter((call) => call.fn === "append_agentic_rectification_turn").length, 0);
+    resetDeliveryTurnGuardForTests();
+  });
+}
+
+test("the exit fill-in still writes once when the reply is only the acknowledgement", async () => {
+  resetDeliveryTurnGuardForTests();
+  markDeliveryTurn({ caseId: CASE_ID, resultId: RESULT_ID, narration: REWRITE });
+  const accounting = gateAccounting("已记录，范围没变。");
+  await persistExhaustionGateTurn({
+    accounting: accounting.client,
+    userId: USER_ID,
+    caseId: CASE_ID,
+    hostNarration: FALLBACK,
+    resultId: RESULT_ID,
+  });
+  assert.equal(accounting.calls.filter((call) => call.fn === "append_agentic_rectification_turn").length, 1);
+  resetDeliveryTurnGuardForTests();
+});
+
+// ---------------------------------------------------------------------------
+// Fix sheet 2026-10-06 (F2): guard only. Widening must never deliver or adopt
+// on the old window's tied candidates, even when the stored snapshot's
+// evidence fingerprint still matches and the rescore after widening fails.
+// (The suspected BUG-1246 did not hold: widen/block choices decide with no
+// inference state, so they always collect; this pins that.)
+// ---------------------------------------------------------------------------
+
+const NARROW = { start_time: RANGE_START, end_time: RANGE_END };
+const WIDE = { start_time: "04:15", end_time: "05:45" };
+
+function widenFocus() {
+  const frame = buildWidenWindowFrame({ reportedTime: TIME_B, currentWindow: NARROW });
+  assert.ok(frame);
+  return activeFocusFixture({
+    intent: WIDEN_WINDOW_INTENT,
+    targetDomain: null,
+    targetKind: null,
+    questionId: "widen_window:search_range:holdout",
+    expectedAnswerSchema: {
+      choice: {
+        prompt: frame.prompt,
+        option_a: frame.option_a_hint,
+        option_b: frame.option_b_hint,
+        option_c: frame.neither_label,
+        option_d: frame.unsure_label,
+        options: [
+          { key: "A", label: frame.option_a_hint, answer_class: frame.option_a_answer_class, role: "primary" },
+          { key: "B", label: frame.option_b_hint, answer_class: frame.option_b_answer_class, role: "primary" },
+          { key: "C", label: frame.neither_label, answer_class: frame.option_c_answer_class, role: "primary" },
+          { key: "D", label: frame.unsure_label, answer_class: frame.option_d_answer_class, role: "primary" },
+        ],
+      },
+      choice_kind: WIDEN_WINDOW_KIND,
+      scoring: false,
+      widen_windows: {
+        A: { radius: 30, ...WIDE },
+        B: { radius: 60, start_time: "04:00", end_time: "06:00" },
+      },
+    },
+  });
+}
+
+test("answering widen A with a failed rescore never delivers or adopts on the old window's candidates", async () => {
+  const draft = parsed(false, true, null);
+  const fingerprint = evidenceLedgerFingerprint(draft.evidence);
+  // Control: the same stored snapshot, read as current, is a tied delivery that may be adopted.
+  const control = publicDecisionFields(decideFromDossier(
+    parsed(false, true, fingerprint),
+    decisionInputsFromSnapshot(BIRTH, { currentEvidenceFingerprint: fingerprint }),
+  ));
+  assert.equal(control.can_adopt, true);
+
+  const current = { value: rawDossier(false, true, fingerprint, { candidateRange: NARROW, activeFocus: widenFocus() }) };
+  const accounting = fakeAccounting({
+    [CASE_DOSSIER_RPC]: () => current.value,
+    get_agentic_rectification_case_compute: () => ({
+      case_id: CASE_ID,
+      skill_version: "9.0.0",
+      baseline_profile_fingerprint: "a".repeat(64),
+      baseline_birth_snapshot: {
+        birth_date: BIRTH,
+        latitude: 31.23,
+        longitude: 121.47,
+        timezone_id: "Asia/Shanghai",
+        timezone_offset: 8,
+        birth_time_source: "approximate",
+        reported_birth_time: TIME_B,
+      },
+      candidate_range: (current.value.case as { candidate_range: unknown }).candidate_range,
+    }),
+    append_agentic_rectification_turn: () => ({ turn_id: "44444444-4444-4444-8444-444444444444", idempotent: false }),
+    apply_agentic_rectification_choice_action: (_fn: string, args: Record<string, unknown>) => ({
+      action_id: args.p_action_id,
+      status: "applied",
+      idempotent: false,
+      question_id: args.p_question_id,
+      option_id: args.p_option_id,
+      revision: 1,
+      narration: args.p_narration,
+      focus_status: args.p_focus_status,
+    }),
+    widen_agentic_rectification_dated_window: () => {
+      current.value = rawDossier(false, true, fingerprint, { candidateRange: WIDE });
+      return { case_id: CASE_ID, stage: "minute", candidate_range: WIDE };
+    },
+  });
+  const savedEnv = {
+    algorithm: process.env.RECTIFICATION_ALGORITHM_VERSION,
+    policy: process.env.RECTIFICATION_DECISION_POLICY_VERSION,
+  };
+  process.env.RECTIFICATION_ALGORITHM_VERSION = ALGO;
+  process.env.RECTIFICATION_DECISION_POLICY_VERSION = POLICY;
+  let receipt: Awaited<ReturnType<typeof applyRectificationChoice>>;
+  try {
+    receipt = await applyRectificationChoice(accounting.client, {
+      userId: USER_ID,
+      caseId: CASE_ID,
+      sessionId: SESSION_ID,
+      actionId: "aaaaaaaa-aaaa-4aaa-8aaa-aaaaaaaaab46",
+      action: CHOICE_ACTION,
+      focusId: FOCUS_ID,
+      optionId: "A",
+      expectedRevision: 0,
+    });
+  } finally {
+    for (const [key, value] of [
+      ["RECTIFICATION_ALGORITHM_VERSION", savedEnv.algorithm],
+      ["RECTIFICATION_DECISION_POLICY_VERSION", savedEnv.policy],
+    ] as const) {
+      if (value === undefined) delete process.env[key];
+      else process.env[key] = value;
+    }
+  }
+  assert.ok(accounting.calls.some((call) => call.fn === "widen_agentic_rectification_dated_window"));
+  assert.equal(receipt.snapshotCurrent, false);
+  assert.equal(receipt.nextAction.can_adopt, false);
+  assert.notEqual(receipt.nextAction.type, "ready_to_adopt");
+  assert.notEqual(receipt.nextAction.type, "complete_with_range");
+});