Knuth-Plass: keep only the best active node per line and fitness class

PanoramicData.Render.ReferenceGenerator/PanoramicData.Render.ReferenceGenerator.csproj

@@ -8,6 +8,12 @@
     <UseWindowsForms>false</UseWindowsForms>
     <NeutralResourcesLanguage>en</NeutralResourcesLanguage>
   </PropertyGroup>
+  <ItemGroup>
+    <!-- Drives Word through COM, so it only runs on Windows; this keeps CA1416 satisfied on the net10.0 target. -->
+    <AssemblyAttribute Include="System.Runtime.Versioning.SupportedOSPlatformAttribute">
+      <_Parameter1>windows</_Parameter1>
+    </AssemblyAttribute>
+  </ItemGroup>
   <ItemGroup>
     <PackageReference Include="DocumentFormat.OpenXml" />
     <PackageReference Include="Microsoft.Office.Interop.Word" />

PanoramicData.Render.Test/KnuthPlassTests.cs

@@ -427,6 +427,33 @@ public void FindBreaks_LooseFitnessOnMidLine_ProducesMultipleLines()
 		lines.Should().HaveCountGreaterThan(1);
 	}
 
+	[Fact]
+	public async Task FindBreaks_LongWordWrappedParagraph_CompletesQuickly()
+	{
+		// Shaped like RenderCommandEmitter.BuildWrapItems output: space glue stretches 0.5x and shrinks 1/3x.
+		// Before active nodes were deduplicated, a paragraph like this took minutes and millions of nodes.
+		var items = new List<KnuthPlassItem>();
+		for (var word = 0; word < 40; word++)
+		{
+			if (word > 0)
+			{
+				items.Add(new KnuthPlassGlue(100f, 50f, 100f / 3f));
+			}
+
+			items.Add(new KnuthPlassBox(300f + (word % 7 * 60f)));
+		}
+
+		items.Add(new KnuthPlassGlue(0f, float.PositiveInfinity, 0f));
+		items.Add(ForcedBreak());
+
+		var lines = await Task
+			.Run(() => KnuthPlassAlgorithm.FindBreaks(items, 9000f), TestContext.Current.CancellationToken)
+			.WaitAsync(TimeSpan.FromSeconds(10), TestContext.Current.CancellationToken);
+
+		lines.Should().NotBeEmpty();
+		lines[^1].EndIndex.Should().Be(items.Count - 1);
+	}
+
 	private static KnuthPlassPenalty ForcedBreak() =>
 		new(0f, KnuthPlassPenalty.NegativeInfinity);
 }

PanoramicData.Render/KnuthPlassAlgorithm.cs

@@ -195,12 +195,20 @@ public static IReadOnlyList<KnuthPlassLine> FindBreaks(
 				activeNodes.RemoveAt(toDeactivate[d]);
 			}
 
-			// Add new active nodes (keep only the best per fitness class for this position)
+			// Add new active nodes, keeping only the best per (line, fitness class): without this the active set doubles at
+			// every breakpoint while lines are too loose. On a tie the later candidate wins, matching the previous output.
+			var bestByLineAndFitness = new Dictionary<(int Line, int FitnessClass), BreakNode>();
 			foreach (var node in toActivate)
 			{
-				activeNodes.Add(node);
+				var key = (node.Line, node.FitnessClass);
+				if (!bestByLineAndFitness.TryGetValue(key, out var existing) || node.TotalDemerits <= existing.TotalDemerits)
+				{
+					bestByLineAndFitness[key] = node;
+				}
 			}
 
+			activeNodes.AddRange(bestByLineAndFitness.Values);
+
 			// If a forced break, deactivate all preceding active nodes
 			if (isForcedBreakHere)
 			{