Enhance table title handling and improve heading detection logic

- Added logic to set `table_title` from `caption` if it is under 160 characters and not already set.
- Updated `_blocks_from_content_list` to assign `table_title` based on `caption` length.
- Introduced new regex patterns for better detection of TOC entries and noise.
- Enhanced heading detection to differentiate between numbered instructions and actual headings.
- Added tests to verify that table captions are correctly assigned as titles and that numbered instructions are treated as body text.

backend/tests/test_heading_layout_multimodal.py

@@ -59,6 +59,100 @@ def table(page: int = 1, y: int = 220) -> Block:
 
 
 class HeadingLayoutMultimodalTest(unittest.TestCase):
+    def test_numbered_instructions_remain_body_text(self) -> None:
+        chunks = split_pdf_semantic(
+            [
+                h("1 Data Platform", level=1),
+                p("1. 用户点击发布则会显示对应的数据发布弹窗", y=140),
+                p("(1) 发布路径：下拉选择菜单，只可单选", y=170),
+                h("1.1 发布管理", level=2, y=220),
+                p("发布管理正文。", y=250),
+            ],
+            SplitConfig(mode=SplitMode.DEFAULT, max_chunk_size=2000),
+        )
+
+        self.assertEqual([chunk.meta["heading"] for chunk in chunks], ["1 Data Platform", "1.1 发布管理"])
+        self.assertIn("1. 用户点击发布", chunks[0].content)
+        self.assertNotIn("# 1. 用户点击发布", chunks[0].content)
+
+    def test_numbered_option_description_is_not_a_heading(self) -> None:
+        option = h(
+            "10.4 Default: This option enables you to set default Qty, Account and/or Best Price.",
+            level=2,
+            y=160,
+        )
+        option.meta["source_heading"] = True
+        chunks = split_pdf_semantic(
+            [h("10 Option settings", level=1), option, p("Following option details.", y=200)],
+            SplitConfig(mode=SplitMode.DEFAULT, max_chunk_size=2000),
+        )
+
+        self.assertEqual(len(chunks), 1)
+        self.assertEqual(chunks[0].meta["heading"], "10 Option settings")
+        self.assertIn("10.4 Default: This option enables", chunks[0].content)
+        self.assertNotIn("## 10.4 Default", chunks[0].content)
+
+    def test_pure_numeric_heading_keeps_its_real_depth(self) -> None:
+        numeric = h("1.1.1", level=2, page=2)
+        numeric.meta["source_heading"] = True
+        next_section = h("1.2 ACCOUNT STATUS CODE MASTER", level=2, page=5)
+        next_section.meta["source_heading"] = True
+        chunks = split_pdf_semantic(
+            [numeric, table(page=2), next_section, p("Status body.", page=5)],
+            SplitConfig(mode=SplitMode.DEFAULT, max_chunk_size=4000),
+        )
+
+        self.assertEqual([chunk.meta["heading"] for chunk in chunks], ["1.1.1", "1.2 ACCOUNT STATUS CODE MASTER"])
+        self.assertNotIn("1.1.1", chunks[1].content)
+        self.assertEqual(chunks[1].meta["pages"], [5])
+
+    def test_clock_text_is_not_a_heading_or_ancestor(self) -> None:
+        clock = h("15:18:53", level=2, page=8)
+        clock.meta["source_heading"] = True
+        chunks = split_pdf_semantic(
+            [clock, h("第三节 买入/卖出序", level=2, page=9), p("Useful body.", page=9)],
+            SplitConfig(mode=SplitMode.DEFAULT, max_chunk_size=2000),
+        )
+
+        self.assertEqual(len(chunks), 1)
+        self.assertEqual(chunks[0].meta["heading"], "第三节 买入/卖出序")
+        self.assertNotIn("15:18:53", chunks[0].content)
+        self.assertEqual(chunks[0].meta["pages"], [9])
+
+    def test_parent_heading_is_metadata_not_repeated_page_block(self) -> None:
+        parent = h("1 Parent", level=1, page=1)
+        parent.meta["source_heading"] = True
+        child = h("1.1 Child", level=2, page=2)
+        child.meta["source_heading"] = True
+        chunks = split_pdf_semantic(
+            [parent, child, p("Child body.", page=2)],
+            SplitConfig(mode=SplitMode.DEFAULT, max_chunk_size=2000),
+        )
+
+        self.assertEqual(len(chunks), 1)
+        self.assertEqual(chunks[0].meta["heading_path"], ["1 Parent", "1.1 Child"])
+        self.assertNotIn("# 1 Parent", chunks[0].content)
+        self.assertEqual(chunks[0].meta["pages"], [2])
+
+    def test_oversized_table_keeps_nearest_heading_and_intro(self) -> None:
+        big_table = table(page=1, y=180)
+        big_table.markdown += "\n" + "| value | description |\n" * 80
+        chunks = split_pdf_semantic(
+            [
+                h("2 Settings", level=1),
+                p("The fields are listed below.", y=140),
+                big_table,
+                p("Following explanation " + "x" * 180, y=340),
+                p("Final paragraph " + "y" * 180, y=380),
+            ],
+            SplitConfig(mode=SplitMode.DEFAULT, max_chunk_size=260),
+        )
+
+        table_chunks = [chunk for chunk in chunks if "table" in chunk.block_types]
+        self.assertEqual(len(table_chunks), 1)
+        self.assertIn("2 Settings", table_chunks[0].content)
+        self.assertIn("The fields are listed below.", table_chunks[0].content)
+        self.assertFalse(any(chunk.block_types and all(t == "heading" for t in chunk.block_types) for chunk in chunks))
     def test_numbered_headings_create_same_level_boundaries(self) -> None:
         chunks = split_pdf_semantic(
             [