diff --git a/docs/task4-compatibility-boundary-evidence.md b/docs/task4-compatibility-boundary-evidence.md
new file mode 100644
index 0000000..b751645
--- /dev/null
+++ b/docs/task4-compatibility-boundary-evidence.md
@@ -0,0 +1,21 @@
+# 任务四：提交失败恢复与可观测性证据
+
+## 问题/目标
+
+`POST /api/submit` 的旧调用仍使用 `code`、`assignment_id`，默认语言仍为 `cpp`。同步评估器抛出异常时，提交已经以 `pending` 写入数据库；目标是在不改变旧响应契约的前提下，保证记录可恢复为 `failed`，并留下可定位日志。
+
+## 复现与证据
+
+在测试中用 `patch('routes.api.evaluate_cpp_code', side_effect=RuntimeError(...))` 调用真实 `/api/submit` 路由。修复前异常分支没有回滚会话，也没有保存异常原因；这会让失败记录缺少恢复信息。测试验证 HTTP 仍为 `500`，同时检查数据库状态和应用日志。
+
+## 计划改动
+
+只修改 `routes/api.py:submit_code` 的同步评估异常分支：先 `db.session.rollback()`，再将同一提交标记为 `failed` 并保存反馈，最后用 `current_app.logger.exception` 记录 `submission_id`。不改请求字段、成功响应、错误状态码、队列接口或数据库结构。
+
+## 学习总结
+
+人工排查沿着 `pending -> evaluated/failed` 状态路径追踪，并把“评估器异常后是否可恢复”写成可证伪假设。拒绝统一错误码、幂等键和异步队列重构等过度设计，因为它们会改变公开契约，不能由当前故障证明必要。
+
+## 验证结果
+
+`python -m pytest tests/test_task4_submission_failure_recovery.py -q`：`1 passed`。回归测试覆盖旧请求字段、500 响应、`failed` 状态、异常反馈和带提交 ID 的应用日志。
diff --git a/routes/api.py b/routes/api.py
index 1aee78a..b652b5a 100644
--- a/routes/api.py
+++ b/routes/api.py
@@ -813,10 +813,15 @@ def submit_code():
             )
             
         except Exception as e:
-            current_app.logger.exception('评估代码失败')
-            
+            # 评估可能在事务中途失败，先回滚会话再保存可恢复的失败状态。
+            db.session.rollback()
             submission.status = 'failed'
+            submission.feedback = f'代码评估失败: {str(e)}'
             db.session.commit()
+            current_app.logger.exception(
+                '评估代码失败，提交已标记为 failed (submission_id=%s)',
+                submission.id,
+            )
             
             return error_response("代码评估失败，请稍后重试", 500)
             
diff --git a/tests/test_task4_submission_failure_recovery.py b/tests/test_task4_submission_failure_recovery.py
new file mode 100644
index 0000000..c99d75d
--- /dev/null
+++ b/tests/test_task4_submission_failure_recovery.py
@@ -0,0 +1,57 @@
+import os
+import tempfile
+from unittest.mock import patch
+
+from app import create_app
+from models import Assignment, Submission, User, db
+
+
+def test_sync_submission_failure_is_persisted_and_observable(caplog):
+    fd, database_path = tempfile.mkstemp()
+    app = create_app('testing')
+    app.config.update(
+        SQLALCHEMY_DATABASE_URI=f'sqlite:///{database_path}',
+        TESTING=True,
+        SUBMISSION_EVALUATION_QUEUE_BACKEND='thread',
+    )
+    client = app.test_client()
+    try:
+        with app.app_context():
+            db.drop_all()
+            db.create_all()
+            student = User(
+                student_id='task4-student',
+                username='task4-student',
+                usertype='学生',
+                full_name='任务四学生',
+            )
+            student.password = 'password'
+            db.session.add(student)
+            assignment = Assignment(title='任务四作业', creator_id=student.student_id)
+            db.session.add(assignment)
+            db.session.commit()
+            assignment_id = assignment.id
+
+        client.post('/login', data={'username': 'task4-student', 'password': 'password'})
+        with patch(
+            'routes.api.evaluate_cpp_code',
+            side_effect=RuntimeError('temporary evaluator outage'),
+        ):
+            response = client.post(
+                '/api/submit',
+                json={'code': 'int main() { return 0; }', 'assignment_id': assignment_id},
+            )
+
+        assert response.status_code == 500
+        assert response.get_json()['success'] is False
+        with app.app_context():
+            submission = Submission.query.one()
+            assert submission.status == 'failed'
+            assert 'temporary evaluator outage' in submission.feedback
+        assert 'submission_id=' in caplog.text
+    finally:
+        with app.app_context():
+            db.session.remove()
+            db.drop_all()
+        os.close(fd)
+        os.unlink(database_path)