Keep top-of-column section headings from being dropped as running headers

The raw-line running-furniture filter treated the whole top ~25% band as header candidates and removed repeating lines before block building, so a "SECTION 1/2/4" heading at the top of a column on a recurring page layout (us_constitution p5/p7/p8/p13/p15) was deleted as a phantom running header -- leaving no element and no bounding box. Tighten the line-level header band to the top 10% (LineHeaderBottomRatio = 0.90); the broader element-level repeating-furniture detector still runs after classification as the real backstop. Adds a regression test asserting those section headings keep positive-area bounding boxes.

Author: diluculo

src/PdfStruct.Tests/UsConstitutionFixtureTests.cs

@@ -87,6 +87,24 @@ public void AllArticlesAndAmendmentsAreExtracted()
             "First Article reference must appear before first Amendment reference in document order.");
     }
 
+    /// <summary>
+    /// Section headings at the top of a page or column are content, not running
+    /// furniture. They must survive with usable bounding boxes so JSON consumers
+    /// can cite the structural boundary directly.
+    /// </summary>
+    [Fact]
+    public void TopOfColumnSectionHeadingsKeepBoundingBoxes()
+    {
+        var path = FixturePath("us_constitution.pdf");
+        var result = new PdfStructParser().Parse(path);
+
+        AssertHasBox(result.Document.Kids, page: 5, text: "SECTION. 1");
+        AssertHasBox(result.Document.Kids, page: 7, text: "SECTION. 1");
+        AssertHasBox(result.Document.Kids, page: 8, text: "SECTION. 1");
+        AssertHasBox(result.Document.Kids, page: 13, text: "SECTION 2", minLeft: 300);
+        AssertHasBox(result.Document.Kids, page: 15, text: "SECTION 4", minLeft: 300);
+    }
+
     /// <summary>Extracts the text content of any text-bearing element, or empty for non-text element types.</summary>
     private static string GetText(ContentElement element) => element switch
     {
@@ -96,6 +114,18 @@ public void AllArticlesAndAmendmentsAreExtracted()
         _ => string.Empty
     };
 
+    private static void AssertHasBox(IEnumerable<ContentElement> elements, int page, string text, double minLeft = 0)
+    {
+        var element = elements.FirstOrDefault(e =>
+            e.PageNumber == page
+            && e.BoundingBox.Left >= minLeft
+            && string.Equals(GetText(e).Trim(), text, StringComparison.Ordinal));
+
+        Assert.NotNull(element);
+        Assert.True(element.BoundingBox.Width > 0 && element.BoundingBox.Height > 0,
+            $"{text} on page {page} must have a positive-area bounding box.");
+    }
+
     /// <summary>Parses a Roman numeral (I..MMMM) into its integer value. Throws on invalid input.</summary>
     private static int ParseRoman(string roman)
     {

src/PdfStruct/PdfStructParser.cs

@@ -90,6 +90,13 @@ public sealed class PdfStructParser
         @"^\s*(자료|출처|source|data source|note|notes)\s*[:：]|^\s*https?://|^\s*www\.|doi\.org|^\s*doi\s*:",
         RegexOptions.IgnoreCase | RegexOptions.Compiled);
 
+    /// <summary>
+    /// Raw line filtering runs before paragraph grouping and heading
+    /// classification, so header candidates must be close to the physical page
+    /// edge. The broader element-level detector still runs after classification.
+    /// </summary>
+    private const double LineHeaderBottomRatio = 0.90;
+
     /// <summary>Initializes with default options.</summary>
     public PdfStructParser() : this(new PdfStructOptions()) { }
 
@@ -2029,9 +2036,9 @@ private static Dictionary<int, IReadOnlyList<TextLineBlock>> FilterRunningFurnit
     /// <summary>
     /// Classifies a line's bounding box as belonging to the page's header,
     /// footer, or side-furniture band, or returns <c>null</c> when the box
-    /// sits in the body region. Header and footer bands are defined by Y
-    /// ratios; side bands require a narrow-and-tall box hugging the left or
-    /// right page edge.
+    /// sits in the body region. The line-level header band is deliberately
+    /// tighter than the post-classification element detector because this pass
+    /// cannot yet distinguish content headings from page furniture.
     /// </summary>
     private static RunningFurnitureBand? ClassifyRunningFurnitureBand(Models.BoundingBox bbox, PageGeometry pageGeometry)
     {
@@ -2042,7 +2049,7 @@ private static Dictionary<int, IReadOnlyList<TextLineBlock>> FilterRunningFurnit
         var topRatio = bbox.Top / pageGeometry.Height;
         if (topRatio < RunningFurnitureDetector.FooterBandBottomRatio)
             return RunningFurnitureBand.Footer;
-        if (bottomRatio > 1.0 - RunningFurnitureDetector.HeaderBandTopRatio)
+        if (bottomRatio > LineHeaderBottomRatio)
             return RunningFurnitureBand.Header;
 
         var nearLeftOrRightEdge = bbox.Right <= pageGeometry.Width * 0.12