fix(reflection): reject incomplete or coerced Judge decisions

docs/architecture.md

@@ -669,7 +669,9 @@ flowchart LR
 不是 Judge 判定通过率；例如正确识别出数字错误的负向夹具也算通过。
 L3 每例和 level `totals` 新增 `judgeAttempted`（实际发起调用，包括异常）和
 `judgeCompleted`（取得可解析且未降级的判定，包括负向判定）；每例为布尔值，汇总为数量。
-这里的“完成”沿用生产反思解析器的接受规则，不额外执行完整字段或类型校验。
+有效判定要求 `answersQuestion`、`consistentWithData`、`needsReplan` 三项均为 JSON 布尔值；
+缺项或使用字符串等其他类型均降级并保留待判定状态，不触发 Judge 策略重规划；独立的数字检查和有数据拒答检查仍生效。
+这里的“完成”沿用生产反思解析器的接受规则；`reason`、`hint` 仍为可选说明，额外属性不影响判定。
 无数据而跳过策略检查的用例不计为尝试、完成或待判定。`judgePending` 表示仍需有效判定的降级用例。
 兼容旧报告消费者，`judgeAvailable` 和 `totals.judgeRan` 保留原有 `reflection.judgeUsed` 语义：
 返回坏 JSON 仍计入，调用抛异常不计入，因此它们不能表示有效判定已完成。看板使用 `judgeCompleted`。

src/reflection.js

@@ -272,17 +272,17 @@ function parseJudgeVerdict(text) {
   }
   try {
     const obj = JSON.parse(match[0]);
-    // Require at least one genuine verdict field so an unrelated JSON payload
-    // (e.g. a presentation contract) is never mistaken for a verdict.
-    const hasVerdictField = ['answersQuestion', 'consistentWithData', 'needsReplan']
-      .some((key) => typeof obj[key] === 'boolean');
-    if (!hasVerdictField) {
+    // Require complete boolean decisions; partial or coerced verdicts cannot
+    // safely determine whether an answer passes or needs a strategy replan.
+    const hasValidDecisions = ['answersQuestion', 'consistentWithData', 'needsReplan']
+      .every((key) => typeof obj[key] === 'boolean');
+    if (!hasValidDecisions) {
       return null;
     }
     return {
-      answersQuestion: Boolean(obj.answersQuestion),
-      consistentWithData: Boolean(obj.consistentWithData),
-      needsReplan: Boolean(obj.needsReplan),
+      answersQuestion: obj.answersQuestion,
+      consistentWithData: obj.consistentWithData,
+      needsReplan: obj.needsReplan,
       reason: String(obj.reason ?? ''),
       hint: String(obj.hint ?? ''),
     };

tests/evalJudge.test.mjs

@@ -15,6 +15,9 @@ for (const { name, judge, attempted, legacyRan } of [
   { name: 'empty', judge: async () => '', attempted: 2, legacyRan: 2 },
   { name: 'null response', judge: async () => null, attempted: 2, legacyRan: 2 },
   { name: 'unrelated JSON', judge: async () => '{}', attempted: 2, legacyRan: 2 },
+  { name: 'partial verdict', judge: async () => '{"answersQuestion":true}', attempted: 2, legacyRan: 2 },
+  { name: 'string consistency', judge: async () => '{"answersQuestion":true,"consistentWithData":"false","needsReplan":false}', attempted: 2, legacyRan: 2 },
+  { name: 'string replan', judge: async () => '{"answersQuestion":true,"consistentWithData":true,"needsReplan":"false"}', attempted: 2, legacyRan: 2 },
 ]) {
   test(`L3 keeps ${name} judgments pending while retaining deterministic fixture checks`, async () => {
     const report = await runLevel('L3', { judge });

tests/reflection.test.mjs

@@ -107,6 +107,90 @@ test('③ strategy sound uses judge verdict (needsReplan triggers hint)', async
   assert.equal(r.hint, '应改按 customer_name 维度拆分而非整体求和');
 });
 
+for (const field of ['answersQuestion', 'consistentWithData', 'needsReplan']) {
+  for (const [name, value] of [
+    ['missing', undefined], ['null', null], ['string true', 'true'],
+    ['string false', 'false'], ['number', 1], ['object', {}], ['array', []],
+  ]) {
+    test(`③ strategy sound degrades without replanning when ${field} is ${name}`, async () => {
+      const verdict = { answersQuestion: true, consistentWithData: true, needsReplan: false };
+      verdict[field] = value;
+      const r = await checkStrategySound({
+        question: '本月销售额', answer: '本月销售额为123.45万元。', result: RESULT,
+        judge: async () => JSON.stringify(verdict),
+      });
+      assert.equal(r.degraded, true);
+      assert.equal(r.judgeUsed, true);
+      assert.equal(r.pass, true);
+      assert.equal(r.needsReplan, false);
+      assert.equal(r.hint, '');
+      assert.equal(r.reason, '');
+    });
+  }
+}
+
+for (const { name, text, pass, needsReplan, reason, hint } of [
+  {
+    name: 'positive without optional text',
+    text: '{"answersQuestion":true,"consistentWithData":true,"needsReplan":false}',
+    pass: true, needsReplan: false, reason: '', hint: '',
+  },
+  {
+    name: 'negative without optional text',
+    text: '{"answersQuestion":false,"consistentWithData":false,"needsReplan":true}',
+    pass: false, needsReplan: true, reason: '', hint: '',
+  },
+  {
+    name: 'fenced JSON with prose, optional text and extra properties',
+    text: 'Verdict follows:\n```json\n{"answersQuestion":true,"consistentWithData":false,"needsReplan":true,"reason":"数字不符","hint":"重查销售额","extra":"allowed"}\n```\nDone.',
+    pass: false, needsReplan: true, reason: '数字不符', hint: '重查销售额',
+  },
+  {
+    name: 'optional text coercion',
+    text: '{"answersQuestion":true,"consistentWithData":true,"needsReplan":false,"reason":123,"hint":null}',
+    pass: true, needsReplan: false, reason: '123', hint: '',
+  },
+]) {
+  test(`③ strategy sound accepts a complete boolean verdict: ${name}`, async () => {
+    const r = await checkStrategySound({
+      question: '本月销售额', answer: '本月销售额为123.45万元。', result: RESULT,
+      judge: async () => text,
+    });
+    assert.equal(r.degraded, false);
+    assert.equal(r.judgeUsed, true);
+    assert.equal(r.pass, pass);
+    assert.equal(r.needsReplan, needsReplan);
+    assert.equal(r.reason, reason);
+    assert.equal(r.hint, hint);
+  });
+}
+
+test('reflectOnAnswer: malformed judge retains numeric failure without triggering a strategy replan', async () => {
+  const r = await reflectOnAnswer({
+    question: '本月销售额', answer: '本月总销售额高达 1.23 亿元。', result: RESULT,
+    judge: async () => '{"answersQuestion":true,"consistentWithData":true,"needsReplan":"false"}',
+  });
+  assert.equal(r.degraded, true);
+  assert.equal(r.judgeUsed, true);
+  assert.equal(r.passed, false);
+  assert.equal(r.checks.find((c) => c.name === 'numeric_consistency').pass, false);
+  assert.equal(r.needsReplan, false);
+  assert.equal(r.hint, null);
+});
+
+test('reflectOnAnswer: malformed judge retains refusal-with-data replanning', async () => {
+  const r = await reflectOnAnswer({
+    question: '本月销售额', answer: '抱歉，我无法确定您要查询的口径。', result: RESULT,
+    judge: async () => '{"answersQuestion":true}',
+  });
+  assert.equal(r.degraded, true);
+  assert.equal(r.judgeUsed, true);
+  assert.equal(r.passed, false);
+  assert.equal(r.checks.find((c) => c.name === 'answerable').refusalWithData, true);
+  assert.equal(r.needsReplan, true);
+  assert.ok(r.hint.includes('不要反问或歉意式拒答'));
+});
+
 test('reflectOnAnswer: clean answer passes without replan (no judge)', async () => {
   const r = await reflectOnAnswer({
     question: '本月销售额前10客户',