Improve LLM rewrite classification pipeline

tests/test_markdown_rendering.py

@@ -27,7 +27,7 @@ class MarkdownRenderingTests(unittest.TestCase):
 
         md, _ = assemble_markdown(items, guide)
 
-        self.assertIn("## 导览", md)
+        self.assertNotIn("## 导览", md)
         self.assertIn("## 模型与能力", md)
         self.assertIn("[OpenAI：Blog ↗](https://openai.com/blog/test)", md)
         self.assertNotIn("> >", md)

tests/test_stage0_to_5_pipeline.py

@@ -37,6 +37,7 @@ class Stage0To5PipelineTests(unittest.TestCase):
                             "id": entry["id"],
                             "title": entry["title_raw"],
                             "summary": entry["summary_raw"],
+                            "section": "模型与能力" if "GPT-5" in entry["title_raw"] else "公司与资本",
                             "flags": [],
                         }
                         for entry in payload["items"]

tests/test_stage0_to_7_pipeline.py

@@ -66,7 +66,7 @@ class Stage0To7PipelineTests(unittest.TestCase):
             guide_llm_call=guide_llm_call,
         )
 
-        self.assertIn("## 导览", result["markdown"])
+        self.assertNotIn("## 导览", result["markdown"])
         self.assertIn("## 模型与能力", result["markdown"])
         self.assertIn("## 今日脉络", result["markdown"])
         self.assertEqual(result["reports"]["stage7"]["blocking_errors"], [])

tests/test_stage4_rewrite.py

@@ -48,6 +48,31 @@ class Stage4RewriteTests(unittest.TestCase):
         self.assertEqual(report["rewritten_count"], 1)
         self.assertEqual(report["fallback_count"], 0)
 
+    def test_rewrite_items_accepts_llm_section_classification(self):
+        items = [news_item("a")]
+
+        def llm_call(prompt):
+            return json.dumps(
+                {
+                    "rewrites": [
+                        {
+                            "id": "a",
+                            "title": "OpenAI 发布 GPT-5 API",
+                            "summary": "OpenAI 发布 GPT-5 API，延迟表现更好。",
+                            "section": "模型与能力",
+                            "confidence": 0.92,
+                            "flags": [],
+                        }
+                    ]
+                },
+                ensure_ascii=False,
+            )
+
+        rewritten, report = rewrite_items(items, llm_call=llm_call, batch_size=10)
+
+        self.assertEqual(rewritten[0].section, "模型与能力")
+        self.assertEqual(report["llm_section_count"], 1)
+
     def test_rewrite_items_falls_back_when_llm_fails(self):
         items = [news_item("a")]
 

tests/test_stage5_classify.py

@@ -45,6 +45,33 @@ class Stage5ClassifyTests(unittest.TestCase):
         self.assertEqual(by_id["b"].section, "开发与基础设施")
         self.assertEqual(report["rule_classified"], 2)
 
+    def test_classify_prefers_valid_llm_section_from_rewrite_stage(self):
+        item = news_item(
+            "a",
+            "API 发布",
+            summary="这其实是一个面向开发者的基础设施能力更新。",
+            section_hint="产品发布/更新",
+        )
+        item.section = "开发与基础设施"
+
+        classified, report = classify_and_order_items([item])
+
+        self.assertEqual(classified[0].section, "开发与基础设施")
+        self.assertEqual(report["llm_classified"], 1)
+        self.assertEqual(report["hint_classified"], 0)
+        self.assertEqual(report["rule_classified"], 0)
+
+    def test_classify_falls_back_when_llm_section_is_invalid(self):
+        item = news_item("a", "GPT-5 发布", section_hint="模型发布/更新")
+        item.section = "热点新闻"
+
+        classified, report = classify_and_order_items([item])
+
+        self.assertEqual(classified[0].section, "模型与能力")
+        self.assertEqual(report["llm_classified"], 0)
+        self.assertEqual(report["hint_classified"], 1)
+        self.assertEqual(report["invalid_llm_section_count"], 1)
+
     def test_classify_orders_items_by_local_rank_score_within_sections(self):
         items = [
             news_item("low", "普通模型更新", section_hint="模型发布/更新", source_priority=80),

tests/test_stage7_assemble.py

@@ -45,8 +45,8 @@ class Stage7AssembleTests(unittest.TestCase):
         md, report = assemble_markdown(items, guide)
 
         self.assertTrue(md.startswith("## 引言\n\n> 今天的 AI 行业继续围绕模型、产品和资本展开。"))
-        self.assertIn("## 导览", md)
-        self.assertIn("> 模型和资本两条线都在推进。", md)
+        self.assertNotIn("## 导览", md)
+        self.assertNotIn("> 模型和资本两条线都在推进。", md)
         self.assertIn("## 模型与能力", md)
         self.assertIn("**1. GPT-5 API 发布**", md)
         self.assertIn("**2. Anthropic 提交 IPO 文件**", md)