feat: LLM-generated adaptive follow-up questions

- Grader: LLM outputs follow_up_question targeting uncovered keyPoints
- Remove static followupHints usage in grading flow
- maxFollowUps sourced from question.maxFollowUps (hints.length)
- Clean answerKey: remove followupHints field
- Three-language prompt update with examples and bad examples
- Grader spec: add follow_up_question to mock responses

server/src/assessment/assessment.service.ts

@@ -532,16 +532,16 @@ private async getModel(tenantId: string): Promise<ChatOpenAI> {
                 difficulty: item.difficulty,
                 dimension: item.dimension,
                 basis: item.basis,
+                maxFollowUps: item.followupHints?.length || 0,
               };
             });
 
-            const answerKey: Record<string, { correctAnswer?: string | null; judgment?: string | null; followupHints?: string[] | null }> = {};
+            const answerKey: Record<string, { correctAnswer?: string | null; judgment?: string | null }> = {};
             selectedItems.forEach(item => {
-              if (item.correctAnswer || item.judgment || item.followupHints) {
+              if (item.correctAnswer || item.judgment) {
                 answerKey[item.id] = {
                   correctAnswer: item.correctAnswer,
                   judgment: item.judgment,
-                  followupHints: item.followupHints,
                 };
               }
             });