fix: degrade distractor pool gracefully on malformed VLM response

tests/unit/test_distractor_selector.py

@@ -55,6 +55,14 @@ class _Material:
     source_nodes = ("n1",)
 
 
+@pytest.mark.asyncio
+async def test_generate_pool_degrades_on_malformed_response():
+    from app.question_gen.distractor_selector import _generate_pool
+    vlm = _FakeVLM(['["not", "a", "dict"]'])  # 顶层是 list 不是 dict
+    out = await _generate_pool(vlm, "?", "蒸", _Material(), 4, session_id="s")
+    assert out == []
+
+
 @pytest.mark.asyncio
 async def test_build_grounded_options_happy_path():
     pool = '{"distractors": ["炒", "煮", "炸", "烤"]}'