Keep a numbered run together when a right-aligned marker rolls a digit

A numbered run with right-aligned markers (plos_game references) slides its left edge left by about one digit width when the number rolls 9->10, so the strict-same-left guard in Candidate.CanExtend vetoed the extension: reference 10 — the last item of the left column — escaped the run, kept its "10." marker, and its bold first line was misread as a heading while its wrapped continuation was orphaned. Grant an extra left-edge tolerance of one digit width per added digit when the marker grows and shifts left, so the run rolls through the digit boundary and reference 10 is emitted as a marker paragraph with its continuation absorbed.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: diluculo

src/PdfStruct.Tests/NumberedParagraphFixtureTests.cs

@@ -66,4 +66,31 @@ public void ConstitutionEnumeration_RendersAsMarkerParagraphs()
         Assert.Contains("1.", markers);
         Assert.Contains("9.", markers);
     }
+
+    [Fact]
+    public void RightAlignedMarkers_SurviveDigitRoll_AtColumnBottom()
+    {
+        // plos_game references use right-aligned markers, so "10." sits a digit
+        // width left of "1."…"9.". The run extension must tolerate that slide;
+        // otherwise reference 10 — the last item of the left column — escapes
+        // the run and its bold first line is misread as a heading.
+        var result = new PdfStructParser().Parse(FixturePath("plos_game_based_education.pdf"));
+
+        var markerParagraphs = result.Document.Kids
+            .OfType<ParagraphElement>()
+            .Where(p => p.Marker is not null)
+            .ToList();
+
+        var tenth = markerParagraphs.FirstOrDefault(p => p.Marker == "10.");
+        Assert.NotNull(tenth);
+        Assert.StartsWith("Martos-Cabrera", tenth!.Text.Content);
+        // The wrapped continuation across the column break is absorbed, not orphaned.
+        Assert.Contains("PMID: 32143452", tenth.Text.Content);
+
+        // Reference 10 is not left as a heading.
+        var numberedHeadings = result.Document.Kids
+            .OfType<HeadingElement>()
+            .Where(h => Regex.IsMatch(h.Text.Content, @"^\d+\.\s"));
+        Assert.Empty(numberedHeadings);
+    }
 }

src/PdfStruct/Analysis/ListDetector.cs

@@ -286,6 +286,28 @@ private static void AbsorbTerritories(
 
     private static double SameLeftTolerance(double fontSize) => Math.Max(fontSize, 1.0) / 3.0;
 
+    /// <summary>
+    /// Approximate width of one marker digit as a fraction of the font size.
+    /// When a numbered run rolls a digit (9→10, 99→100) the marker is right-
+    /// aligned on its terminator, so the left edge slides left by about one
+    /// digit width — a legitimate shift that must not break the run's
+    /// strict-left-alignment streak.
+    /// </summary>
+    private const double MarkerDigitWidthRatio = 0.6;
+
+    /// <summary>
+    /// Extra left-edge tolerance (PDF points) granted when <paramref name="label"/>
+    /// carries more printed digits than <paramref name="lastLabel"/> and the new
+    /// line sits to the left of <paramref name="lastLine"/> — the expected slide
+    /// of a right-aligned marker as it rolls to a wider number. Zero otherwise.
+    /// </summary>
+    private static double DigitGrowthLeftTolerance(TextLineBlock line, ListLabel label, TextLineBlock lastLine, ListLabel lastLabel)
+    {
+        var extraDigits = label.DigitText.Length - lastLabel.DigitText.Length;
+        if (extraDigits <= 0 || line.Left >= lastLine.Left) return 0.0;
+        return extraDigits * Math.Max(lastLine.FontSize, 1.0) * MarkerDigitWidthRatio;
+    }
+
     /// <summary>
     /// Maximum distance, in points, a continuation line may sit to the left of
     /// its item's start before it counts as a structural outdent that ends the
@@ -327,7 +349,7 @@ public bool CanExtend(TextLineBlock line, ListLabel label)
             if (last.Label.Terminator != label.Terminator) return false;
             if (label.Number != last.Label.Number + 1) return false;
 
-            var tol = SameLeftTolerance(last.Line.FontSize);
+            var tol = SameLeftTolerance(last.Line.FontSize) + DigitGrowthLeftTolerance(line, label, last.Line, last.Label);
             var diff = Math.Abs(line.Left - last.Line.Left);
             var sameLeft = diff <= tol;
             var nearLeft = diff <= NearLeftToleranceMultiplier * tol;
@@ -341,7 +363,7 @@ public bool CanExtend(TextLineBlock line, ListLabel label)
         public void Append(int lineIndex, TextLineBlock line, ListLabel label, int labelLength)
         {
             var last = _items[^1];
-            var tol = SameLeftTolerance(last.Line.FontSize);
+            var tol = SameLeftTolerance(last.Line.FontSize) + DigitGrowthLeftTolerance(line, label, last.Line, last.Label);
             var sameLeft = Math.Abs(line.Left - last.Line.Left) <= tol;
             if (!sameLeft) _everyTransitionWasStrictSameLeft = false;