feat: judgment-anchored grading and per-question results

- Grader: inject judgment as pass criteria anchor in LLM prompt
- Grader: use followupHints for follow-up direction (not generic text)
- Grader: follow-up limit from followupHints.length instead of hardcoded 2
- Session: correctAnswer/judgment stored in questions, stripped during assessment
- Frontend: per-question results panel with choice ✅/❌ + judgment display

server/src/assessment/assessment.service.ts

@@ -501,6 +501,8 @@ private async getModel(tenantId: string): Promise<ChatOpenAI> {
               questionText: item.questionText,
               questionType: item.questionType,
               options: item.options,
+              correctAnswer: item.correctAnswer,
+              judgment: item.judgment,
               keyPoints: item.keyPoints,
               difficulty: item.difficulty,
               dimension: item.dimension,
@@ -768,7 +770,10 @@ const initialState: Partial<EvaluationState> = {
             }
             await this.sessionRepository.save(session);
 
-            const mappedData: any = this.sanitizeStateForClient({ ...finalData });
+            const mappedData: any = this.sanitizeStateForClient(
+              { ...finalData },
+              session.status !== AssessmentStatus.COMPLETED,
+            );
             mappedData.messages = this.mapMessages(finalData.messages);
             mappedData.feedbackHistory = this.mapMessages(
               finalData.feedbackHistory || [],
@@ -1139,7 +1144,10 @@ const initialState: Partial<EvaluationState> = {
             }
             await this.sessionRepository.save(session);
 
-            const mappedData: any = this.sanitizeStateForClient({ ...finalData });
+            const mappedData: any = this.sanitizeStateForClient(
+              { ...finalData },
+              session.status !== AssessmentStatus.COMPLETED,
+            );
             mappedData.messages = this.mapMessages(finalData.messages);
             mappedData.feedbackHistory = this.mapMessages(
               finalData.feedbackHistory || [],
@@ -1185,7 +1193,10 @@ const initialState: Partial<EvaluationState> = {
       values.feedbackHistory = this.mapMessages(values.feedbackHistory);
     }
 
-    return this.sanitizeStateForClient(values);
+    return this.sanitizeStateForClient(
+      values,
+      session.status !== AssessmentStatus.COMPLETED,
+    );
   }
 
   /**
@@ -1394,14 +1405,19 @@ const initialState: Partial<EvaluationState> = {
   /**
    * Strips sensitive fields before sending state to frontend.
    */
-  private sanitizeStateForClient(data: any): any {
+  private sanitizeStateForClient(data: any, stripAnswers = true): any {
     if (!data) return data;
     const sanitized = { ...data };
-    delete sanitized.questionAnswerKey;
+    if (stripAnswers) {
+      delete sanitized.questionAnswerKey;
+    }
     if (Array.isArray(sanitized.questions)) {
       sanitized.questions = sanitized.questions.map((q: any) => {
-        const { correctAnswer, judgment, followupHints, ...rest } = q;
-        return rest;
+        if (stripAnswers) {
+          const { correctAnswer, judgment, followupHints, ...rest } = q;
+          return rest;
+        }
+        return q;
       });
     }
     return sanitized;

server/src/assessment/graph/nodes/grader.node.ts

@@ -68,10 +68,10 @@ export const graderNode = async (
   }
 
   const isChoice = currentQuestion.questionType === 'MULTIPLE_CHOICE';
+  const expectedAnswer = currentQuestion.correctAnswer;
   const answerKey = (state.questionAnswerKey as any)?.[currentQuestion.id];
 
-  if (isChoice || answerKey?.correctAnswer) {
-    const expectedAnswer = answerKey?.correctAnswer || currentQuestion.correctAnswer;
+  if (isChoice && expectedAnswer) {
     const userAnswer = (lastUserMessage.content as string).trim();
     const isCorrect = userAnswer.toUpperCase() === expectedAnswer?.toUpperCase();
 
@@ -173,12 +173,24 @@ Format your response as JSON:
   "should_follow_up": false
 }`;
 
-  const systemPrompt = isZh
+  let systemPrompt = isZh
     ? systemPromptZh
     : isJa
       ? systemPromptJa
       : systemPromptEn;
 
+  if (currentQuestion.judgment) {
+    const anchorText = isZh
+      ? `\n\n【判定依据（通过标准）】${currentQuestion.judgment}`
+      : isJa
+        ? `\n\n【判定基準（合格基準）】${currentQuestion.judgment}`
+        : `\n\n【Judgment Criteria (Pass Standard)】${currentQuestion.judgment}`;
+    systemPrompt += anchorText;
+  }
+
+  const followupHints: string[] = answerKey?.followupHints || [];
+  const maxFollowUps = followupHints.length > 0 ? followupHints.length : 2;
+
   const userContentText =
     typeof lastUserMessage.content === 'string'
       ? lastUserMessage.content
@@ -212,10 +224,7 @@ Format your response as JSON:
 
     const scoreLabel = isZh ? '得分' : isJa ? 'スコア' : 'Score';
     const feedbackLabel = isZh ? '反馈' : isJa ? 'フィードバック' : 'Feedback';
-
-    const feedbackMessage = new AIMessage(
-      `${scoreLabel}: ${result.score}/10\n\n${feedbackLabel}: ${result.feedback}`,
-    );
+    let enhancedFeedback: string = result.feedback;
 
     const newScores = {
       ...state.scores,
@@ -224,10 +233,6 @@ Format your response as JSON:
 
     let shouldFollowUp = result.should_follow_up === true;
 
-    // Breakout logic:
-    // 1. Max 1 follow-up per question
-    // 2. If score is decent (>= 8), don't follow up
-    // 3. If answer is short "don't know", don't follow up
     const normalizedContent = userContentText.trim().toLowerCase();
     const saysIDontKnow =
       normalizedContent.length < 10 &&
@@ -242,10 +247,20 @@ Format your response as JSON:
         normalizedContent.includes('不明') ||
         normalizedContent.includes('わからない'));
 
-    if (currentFollowUpCount >= 2 || result.score >= 8 || saysIDontKnow) {
+    if (currentFollowUpCount >= maxFollowUps || result.score >= 8 || saysIDontKnow) {
       shouldFollowUp = false;
     }
 
+    if (shouldFollowUp && followupHints.length > 0) {
+      const hint = followupHints[Math.min(currentFollowUpCount, followupHints.length - 1)];
+      const hintLabel = isZh ? '追问方向' : isJa ? '追加の方向性' : 'Follow-up hint';
+      enhancedFeedback = `${result.feedback}\n\n${hintLabel}: ${hint}`;
+    }
+
+    const feedbackMessage = new AIMessage(
+      `${scoreLabel}: ${result.score}/10\n\n${feedbackLabel}: ${enhancedFeedback}`,
+    );
+
     console.log('[GraderNode] Final State decision:', {
       shouldFollowUp,
       nextIndex: shouldFollowUp