perf: improve AI prompt cache reuse

kouyu_english/test/generated_content_test.dart

@@ -57,7 +57,11 @@ void main() {
         () => MockClient((request) async {
           final body = jsonDecode(request.body) as Map<String, dynamic>;
           final messages = body['messages'] as List<dynamic>;
-          prompts.add((messages.single as Map<String, dynamic>)['content']);
+          expect(messages, hasLength(2));
+          expect(messages.first['role'], 'system');
+          expect(messages.last['role'], 'user');
+          prompts.add(messages.first['content'] as String);
+          prompts.add(messages.last['content'] as String);
           return http.Response(
             jsonEncode({
               'choices': [
@@ -73,7 +77,8 @@ void main() {
       );
 
       expect(variant, isNotNull);
-      expect(prompts, hasLength(1));
+      expect(prompts, hasLength(2));
+      final fullPrompt = prompts.join('\n');
       for (final field in [
         'schemaVersion',
         'variantId',
@@ -81,10 +86,10 @@ void main() {
         'prompt',
         'expectedAnswer',
       ]) {
-        expect(prompts.single, contains(field));
+        expect(fullPrompt, contains(field));
       }
       for (final field in ['stimulus', 'acceptedAnswers', 'forbiddenPhrases']) {
-        expect(prompts.single, isNot(contains(field)));
+        expect(fullPrompt, isNot(contains(field)));
       }
     },
   );