fix(rectification): require three answers before uncertainty stop

One "一时说不好" no longer ends the interview. Plateau streaks use the
same sample floor. Delivery copy states why the range stopped.

Co-authored-by: Cursor <cursoragent@cursor.com>

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

@@ -805,6 +805,48 @@ test("deferFollowup scores the choice without persisting the next interview or t
   assert.equal(applied.narrationPersisted, false);
 });
 
+test("one unsure discriminator answer does not adopt and persists the next question", async () => {
+  const accounting = persistChoiceAccounting(twoProbeDossier(), {
+    set_agentic_rectification_conversation_focus: (_fn, args) => ({
+      focus: {
+        id: NEXT_FOCUS_ID,
+        case_id: CASE_ID,
+        question_id: args.p_question_id,
+        intent: args.p_intent,
+        target_evidence_id: args.p_target_evidence_id,
+        target_domain: args.p_target_domain,
+        target_kind: args.p_target_kind,
+        expected_answer_schema: args.p_expected_answer_schema,
+        status: "active",
+        asked_at: "2026-08-28T07:37:50.000Z",
+        resolved_at: null,
+      },
+      idempotent: false,
+    }),
+  });
+  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: "D",
+    expectedRevision: twoProbeInference().revision,
+  });
+  assert.equal(applied.answerClass, "unsure");
+  assert.equal(applied.narration, "已记录。这题先不计分，换一件事问。");
+  assert.notEqual(applied.nextAction.session_outcome, "adopt_representative");
+  assert.equal(applied.nextAction.type, "ask_candidate_discriminator");
+  assert.equal(applied.nextInterviewPersisted, true);
+  const persist = accounting.calls.find((call) => call.fn === "apply_agentic_rectification_choice_action");
+  assert.ok(persist?.args.p_inference);
+  const setFocus = accounting.calls.find((call) => call.fn === "set_agentic_rectification_conversation_focus");
+  assert.ok(setFocus);
+});
+
 test("answering the last discriminator persists a year-locked family collect focus, not a yearless D24 card", async () => {
   const accounting = persistChoiceAccounting(familyCollectDossier(), {
     set_agentic_rectification_conversation_focus: (_fn, args) => ({
@@ -1116,11 +1158,29 @@ test("structured choice narration never persists a fake loading state", () => {
     composeChoiceNarration({ optionId: "A", scoring: true, appliedInference: true }),
     composeChoiceNarration({ optionId: "A", scoring: true, appliedInference: false }),
     composeChoiceNarration({ optionId: "A", scoring: false, appliedInference: false }),
+    composeChoiceNarration({
+      optionId: "D",
+      scoring: true,
+      appliedInference: true,
+      answerClass: "unsure",
+    }),
   ]) {
     assert.doesNotMatch(narration, /正在准备下一步/);
   }
 });
 
+test("an unsure choice keeps applied inference and does not claim scores changed", () => {
+  const narration = composeChoiceNarration({
+    optionId: "D",
+    scoring: true,
+    appliedInference: true,
+    answerClass: "unsure",
+  });
+  // 旧：appliedInference 仍说「更新了候选比较」→ 新：unsure 不计分旁白 → 保留 inference 行仍持久化
+  assert.equal(narration, "已记录。这题先不计分，换一件事问。");
+  assert.doesNotMatch(narration, /更新了候选比较/);
+});
+
 test("the public agent route treats structured choice as a non-model command", () => {
   const route = readFileSync(new URL("../src/app/api/rectification/agent/route.ts", import.meta.url), "utf8");
   const start = route.indexOf("if (isStructuredChoice)");