fix(rectification): persist narration without routing domains

docs/BUG_HISTORY.md

@@ -1459,3 +1459,18 @@
 - 防复发：叙事模型只能影响自然语言表达，不能成为事件持久化和确定性收敛的单点故障；降级输出必须来自已验证 packet。
 - 相关记录：BUG-067、BUG-075、BUG-077
 - 修复版本：本次修复提交
+
+## BUG-079 | 无可用路由领域时 Agent 追问导致 paused 会话返回 503
+
+- 状态：resolved
+- 首次发现：2026-07-26
+- 最近更新：2026-07-26
+- 影响面：生时校正开放叙事、暂停后恢复、历史事件保存、Narrative Agent 到公开 turn 的适配边界
+- 用户现象：完整 synthetic smoke 在暂停并恢复后提交下一条历史事件，技术评分和模型叙事都成功，但接口返回 `service_unavailable`，事件无法保存。
+- 触发条件：技术 packet 的 `suggestedDomains` 为空，而模型自然提出了一个可继续回答的问题。
+- 根因：Narrative Agent 会把模型领域替换成服务端允许的领域；允许集合为空时产生 `evidenceRequest.domains=[]`。叙事校验未拒绝空数组，但公开 turn 合同要求至少一个领域，导致 `turnFromNarrative()` 在持久化前把整轮转换为 503。
+- 修复：在 Narrative Agent 的共享适配边界统一处理 authored 与 legacy 输出；没有任何服务端可验证路由领域时保留自然语言正文，只把可选 `evidenceRequest` 归一化为 `null`。不放宽公开持久化 schema，也不信任模型自报领域。
+- 验证：回归覆盖无路由领域的 authored 输出，并覆盖 `pause → resume → answer` 完整编排路径；确认正文保留、`evidenceRequest=null`、事件正常保存且不返回 503。
+- 防复发：模型自然语言不得成为事件保存的硬依赖；私有路由元数据为空时应省略可选状态，而不是生成违反持久化合同的半合法对象。
+- 相关记录：BUG-075、BUG-078
+- 修复版本：本次修复提交

frontend/src/lib/conversational-rectification/narrative-agent.ts

@@ -168,16 +168,20 @@ function groundedEvidenceDomains(
   return grounded.length > 0 ? grounded : suggested;
 }
 
+function groundedEvidenceRequest(
+  request: z.infer<typeof rectificationNarrativeAuthoredOutputSchema>["evidenceRequest"],
+  packet: RectificationTechnicalPacket,
+) {
+  if (request === null) return null;
+  const domains = groundedEvidenceDomains(request.domains, packet);
+  return domains.length > 0 ? { ...request, domains } : null;
+}
+
 function completeAuthoredOutput(
   output: z.infer<typeof rectificationNarrativeAuthoredOutputSchema>,
   packet: RectificationTechnicalPacket,
 ): RectificationNarrativeModelOutput {
-  const evidenceRequest = output.evidenceRequest === null ? null : {
-    datePrecision: output.evidenceRequest.datePrecision,
-    prompt: output.evidenceRequest.prompt,
-    followUp: output.evidenceRequest.followUp,
-    domains: groundedEvidenceDomains(output.evidenceRequest.domains, packet),
-  };
+  const evidenceRequest = groundedEvidenceRequest(output.evidenceRequest, packet);
   return {
     ...output,
     evidenceRequest,
@@ -206,14 +210,10 @@ function parseModelOutput(
   if (legacy.success) {
     return {
       ...legacy.data,
-      evidenceRequest: legacy.data.evidenceRequest === null ? null : {
-        ...legacy.data.evidenceRequest,
-        // The next conversational topic is authored by the model, but the
-        // scoring-domain allowlist remains server-owned. Do not let a useful
-        // answer fail merely because the model described that topic with a
-        // different internal domain label.
-        domains: groundedEvidenceDomains(legacy.data.evidenceRequest.domains, packet),
-      },
+      // The next conversational topic is authored by the model, but the
+      // scoring-domain allowlist remains server-owned. If no grounded routing
+      // domain exists, keep the prose and omit only the optional follow-up state.
+      evidenceRequest: groundedEvidenceRequest(legacy.data.evidenceRequest, packet),
     };
   }
   return completeAuthoredOutput(rectificationNarrativeAuthoredOutputSchema.parse(parsed), packet);

frontend/tests/conversational-narrative-agent.test.ts

@@ -935,6 +935,28 @@ test("replaces legacy model-selected evidence domains instead of rejecting the a
   assert.deepEqual(result.output.evidenceRequest?.domains, ["relationship", "career"]);
 });
 
+test("keeps authored prose but drops follow-up state when no grounded routing domain exists", async () => {
+  const packet = { ...syntheticTechnicalPacket(), suggestedDomains: [] };
+  const result = await generateRectificationNarrative({
+    phase: "intermediate",
+    packet,
+    generator: generator([{
+      narrative: "这段经历已经记下。你愿意的话，可以继续讲当时发生了什么。",
+      evidenceRequest: {
+        domains: ["career"],
+        datePrecision: "month_preferred",
+        prompt: "当时发生了什么？",
+        followUp: { kind: "new_event", evidenceId: null },
+      },
+    }]),
+  });
+
+  assert.equal(result.attempts, 1);
+  assert.equal(result.fallbackUsed, false);
+  assert.equal(result.output.evidenceRequest, null);
+  assert.match(result.narrative, /继续讲/);
+});
+
 test("uses a fresh second provider request after the first attempt times out", async () => {
   let calls = 0;
   const signals: Array<AbortSignal | undefined> = [];

frontend/tests/conversational-rectification-orchestrator.test.ts

@@ -1641,22 +1641,22 @@ test("the next evidence request moves past a domain the user already answered",
   assert.doesNotMatch(turn.narrative, /下一步[^\n]*重要关系/);
 });
 
-test("a rejected intermediate narrative returns a retryable error without saving a template turn", async () => {
+test("a rejected intermediate narrative falls back without discarding the event", async () => {
   const value = harness({ invalidNarrativeFromGeneration: 1 });
   await start(value, null);
 
-  await assert.rejects(value.service.answer(userId, {
+  const turn = await value.service.answer(userId, {
     type: "answer", caseId: startActionId, actionId: answerActionId,
     turnVersion: 0, answer: "2021年7月开始第一份长期工作",
-  }), (error: unknown) => error instanceof ConversationalRectificationError
-    && error.code === "service_unavailable");
+  });
 
   const stored = value.cases.get(startActionId)?.row;
-  assert.equal(stored?.turnVersion, 0);
+  assert.equal(turn.turnVersion, 1);
+  assert.match(turn.narrative, /按自己的节奏/);
   assert.equal(stored?.privateCandidate.resultId, null);
-  assert.equal(stored?.eventEvidence.length, 0);
-  assert.equal(stored?.validationReceipts.length, 1);
-  assert.equal(value.mutations.filter((mutation) => mutation === "saveTurn").length, 0);
+  assert.equal(stored?.eventEvidence.length, 1);
+  assert.equal(stored?.validationReceipts.length, 2);
+  assert.equal(value.mutations.filter((mutation) => mutation === "saveTurn").length, 1);
 });
 
 test("one through three supported events save and narrate before the fourth accumulated event ranks", async () => {
@@ -1785,6 +1785,39 @@ test("an unanswered suggested domain keeps a plateaued candidate conversational"
   assert.deepEqual(latest?.actions, ["answer", "pause", "abandon"]);
 });
 
+test("answer persists after pause when the packet has no grounded follow-up domain", async () => {
+  const value = harness({
+    packetForEvidenceCount() {
+      return { ...packet(false), suggestedDomains: [] };
+    },
+  });
+  const started = await start(value, null);
+  const paused = await value.service.pause(userId, {
+    type: "pause",
+    caseId: startActionId,
+    actionId: pauseActionId,
+    turnVersion: started.turnVersion,
+  });
+  await value.service.resume(userId, {
+    type: "resume",
+    caseId: startActionId,
+    actionId: resumeActionId,
+    turnVersion: paused.turnVersion,
+  });
+
+  const turn = await value.service.answer(userId, {
+    type: "answer",
+    caseId: startActionId,
+    actionId: answerActionId,
+    turnVersion: paused.turnVersion,
+    answer: "2020年4月进入研究院实习，10月主动辞职",
+  });
+
+  assert.equal(turn.status, "active");
+  assert.equal(turn.evidenceRequest, null);
+  assert.match(turn.narrative, /记下了/);
+});
+
 test("system-only blockers return a bounded result without waiting for another plateau", async () => {
   const value = harness({
     packetForEvidenceCount(count) {
@@ -2034,7 +2067,7 @@ test("regenerate rewrites only the current narrative and preserves evidence, sco
   assert.match(value.narrativePrompts.at(-1) ?? "", /2012年12月正式退学/);
 });
 
-test("a failed regenerate preserves the prior turn, evidence, candidate, and billing", async () => {
+test("a failed regenerate saves a fallback turn without changing evidence, candidate, or billing", async () => {
   const value = harness({ readyAfterEvidenceCount: 99, invalidNarrativeFromGeneration: 2 });
   await start(value, null);
   const answered = await value.service.answer(userId, {
@@ -2046,28 +2079,29 @@ test("a failed regenerate preserves the prior turn, evidence, candidate, and bil
   });
   const storedBefore = value.cases.get(startActionId)?.row;
   assert.ok(storedBefore);
-  const snapshotBefore = structuredClone(storedBefore);
+  const evidenceBefore = structuredClone(storedBefore.eventEvidence);
+  const candidateBefore = structuredClone(storedBefore.privateCandidate);
   const countsBefore = value.counts();
   const saveTurnsBefore = value.mutations.filter((mutation) => mutation === "saveTurn").length;
 
-  await assert.rejects(value.service.regenerate(userId, {
+  const regenerated = await value.service.regenerate(userId, {
     type: "regenerate",
     caseId: startActionId,
     actionId: laterActionId,
     turnVersion: answered.turnVersion,
-  }), (error: unknown) => error instanceof ConversationalRectificationError
-    && error.code === "service_unavailable");
+  });
 
   const storedAfter = value.cases.get(startActionId)?.row;
-  assert.deepEqual(storedAfter, snapshotBefore);
-  assert.equal(storedAfter?.turnVersion, answered.turnVersion);
-  assert.equal(storedAfter?.latestTurn.narrative, answered.narrative);
+  assert.equal(regenerated.turnVersion, answered.turnVersion + 1);
+  assert.match(regenerated.narrative, /按自己的节奏/);
+  assert.deepEqual(storedAfter?.eventEvidence, evidenceBefore);
+  assert.deepEqual(storedAfter?.privateCandidate, candidateBefore);
   assert.equal(value.counts().reserveCount, countsBefore.reserveCount);
   assert.equal(value.counts().releaseCount, countsBefore.releaseCount);
   assert.equal(value.counts().packetBuilds, countsBefore.packetBuilds + 1);
   assert.equal(
     value.mutations.filter((mutation) => mutation === "saveTurn").length,
-    saveTurnsBefore,
+    saveTurnsBefore + 1,
   );
 });