fix(interview): stabilize VAD orchestration and structured LLM output

README.md

@@ -187,6 +187,35 @@ openssl rand -base64 32
 使用场景生成、Realtime、TTS、ASR 或评分能力时，还需在 `.env` 中配置对应厂商凭证。
 完整变量及安全默认值见 [`deploy/env/.env.example`](deploy/env/.env.example)。
 
+### 2.1 本地启用 JD 图片 OCR
+
+Web 的“上传图片”入口以服务端探测结果为准，不需要手动设置浏览器端的开关。后端本地运行时，先准备 Python 3.11、PaddleOCR 依赖和模型：
+
+```bash
+./scripts/prepare-local-ocr.sh
+```
+
+然后从 `backend/unispeaking-server` 启动后端，并把 OCR 路径指向仓库内的本地目录：
+
+```bash
+cd backend/unispeaking-server
+OCR_ENABLED=true \
+OCR_PYTHON_EXECUTABLE="$PWD/../../.local/ocr/venv/bin/python" \
+OCR_RUNNER_PATH="$PWD/src/main/resources/ocr/paddle_ocr_runner.py" \
+OCR_MODEL_DIRECTORY="$PWD/../../.local/ocr/models" \
+MAVEN_REPO_URL=https://maven.aliyun.com/repository/public \
+./mvnw --settings docker/maven/settings.xml spring-boot:run
+```
+
+启动后用浏览器访问 Web，在模拟面试页面选择“上传图片”。也可以用登录后的 JWT 进行接口实测：
+
+```bash
+OCR_ACCESS_TOKEN='登录后 localStorage 中的 unispeaking.accessToken' \
+./scripts/check-local-ocr.sh /absolute/path/to/jd.png
+```
+
+脚本先验证 `/api/interview-scenes/ocr/availability`，再提交图片到 `/prepare-materials`；这样可以区分“服务端未装好 OCR”和“图片上传/材料整理链路失败”。
+
 注意：
 
 - 不要提交真实 `.env`。

backend/unispeaking-server/src/main/java/com/unispeaking/component/report/InterviewReportCoordinator.java

@@ -517,7 +517,14 @@ private String buildLlmPrompt(
 				- VOCABULARY_EXPRESSION (0-100)
 
 				Produce an overall_score (0-100) as your comprehensive judgment across all five
-				dimensions, and a short summary narrative of the candidate's spoken English.
+				dimensions.
+
+				LANGUAGE REQUIREMENT: Write the natural-language values of ALL "evaluation",
+				"advice", and "summary" fields in Simplified Chinese. Do not return those
+				fields in English. Keep standard English linguistic terms, quoted candidate
+				English, and short English examples only when they are necessary as evidence;
+				these exceptions must be embedded in an otherwise Chinese explanation. The JSON
+				property names and score values must remain exactly as specified below.
 
 				Return exactly one JSON object and no Markdown or explanatory prose.
 				The JSON shape must be:

backend/unispeaking-server/src/main/java/com/unispeaking/infrastructure/ai/qwen/QwenLlmProvider.java

@@ -3,6 +3,7 @@
 import com.unispeaking.common.exception.BusinessException;
 import com.unispeaking.provider.AiProviderRegistry;
 import com.unispeaking.provider.LlmProvider;
+import com.unispeaking.provider.LlmResponseFormat;
 import java.io.ByteArrayOutputStream;
 import java.io.IOException;
 import java.net.URI;
@@ -13,6 +14,7 @@
 import java.nio.charset.StandardCharsets;
 import java.time.Duration;
 import java.util.List;
+import java.util.LinkedHashMap;
 import java.util.Map;
 import java.util.Set;
 import java.util.concurrent.CompletableFuture;
@@ -82,26 +84,40 @@ public QwenLlmProvider(
 
 	@Override
 	public String executeLlmTask(String prompt, String token) {
+		return executeLlmTask(prompt, token, LlmResponseFormat.TEXT);
+	}
+
+	@Override
+	public String executeLlmTask(
+			String prompt,
+			String token,
+			LlmResponseFormat responseFormat) {
 		if (apiKey.isBlank()) {
 			throw retryableFailure(
 					"QWEN_LLM_CREDENTIAL_MISSING",
 					"Set DASHSCOPE_API_KEY before calling Qwen LLM");
 		}
-		return callForContent(prompt, apiKey);
+		return callForContent(prompt, apiKey, responseFormat);
 	}
 
-	private String callForContent(String promptValue, String credential) {
+	private String callForContent(
+			String promptValue,
+			String credential,
+			LlmResponseFormat responseFormat) {
 		String prompt = trim(promptValue);
 		if (prompt.isBlank()) {
 			throw nonRetryableFailure("INVALID_LLM_PROMPT", "LLM task prompt is required");
 		}
 		requireHttpsEndpoint();
 
 		try {
-			Map<String, Object> body = Map.of(
-					"model", model,
-					"messages", List.of(Map.of("role", "user", "content", prompt)),
-					"enable_thinking", false);
+			Map<String, Object> body = new LinkedHashMap<>();
+			body.put("model", model);
+			body.put("messages", List.of(Map.of("role", "user", "content", prompt)));
+			body.put("enable_thinking", false);
+			if (responseFormat == LlmResponseFormat.JSON_OBJECT) {
+				body.put("response_format", Map.of("type", "json_object"));
+			}
 			HttpRequest httpRequest = HttpRequest.newBuilder()
 					.uri(endpoint)
 					.timeout(readTimeout)