Optimize per-edge gateway passes: cheap validation before full scoring

Add per-edge node-crossing and shared-lane pre-check before expensive
ComputeScore. Skip final boundary-slot snap in low-wave path (no-op:
violations 4->4). Boundary-slot polish kept (fixes entry-angle).

Layout-only speed regressed from 14s to ~2m due to quality pipeline
additions (boundary-slot polish 49s, detour polish 25s, per-edge
gateway redirect+scoring). This is the tradeoff for zero-violation
artifact quality. Speed optimization is future work.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

src/__Libraries/StellaOps.ElkSharp/ElkEdgeRouterIterative.WinnerRefinement.Hybrid.cs

@@ -249,7 +249,10 @@ internal static partial class ElkEdgeRouterIterative
         // artifact polish ends with NormalizeBoundaryAngles +
         // NormalizeSourceExitAngles, which is the root cause of the
         // boundary-slot violations when snap ran before it.
-        if (current.RetryState.BoundarySlotViolations > 0)
+        // Final boundary-slot snap is expensive (~39s). Skip in the low-wave
+        // speed path since the FinalScore already handles boundary-slot
+        // exclusions. The full-wave path runs it for maximum quality.
+        if (current.RetryState.BoundarySlotViolations > 0 && !preferLowWaveRuntimePolish)
         {
             current = ApplyFinalBoundarySlotPolish(current, nodes, direction, minLineClearance, maxRounds: 1);
             ElkLayoutDiagnostics.LogProgress($"Hybrid winner refinement after final boundary-slot snap: {DescribeSolution(current)}");
@@ -267,27 +270,17 @@ internal static partial class ElkEdgeRouterIterative
             ElkLayoutDiagnostics.LogProgress($"Hybrid winner refinement after gateway diagonal straightening: {DescribeSolution(current)}");
         }
 
-        // Per-edge gateway source face redirect: process each gateway artifact
+        // Per-edge gateway fixes: only run when the gateway artifact polish
-        // edge individually, applying the face redirect and validating that it
+        // left remaining artifacts. Skip the expensive per-edge scoring when
-        // doesn't create new hard-rule violations. Bulk processing (all edges at
+        // artifacts are already clean.
-        // once) creates 19+ boundary-slot violations because the redirect paths
-        // conflict with each other. Per-edge with validation avoids cascading.
         var postArtifactState = EvaluateGatewayArtifacts(current.Edges, nodes, out var postFocus);
         if (!postArtifactState.IsClean && postFocus.Length > 0)
         {
             current = ApplyPerEdgeGatewayFaceRedirect(current, nodes, minLineClearance, postFocus);
+            current = ApplyPerEdgeGatewayScoringFix(current, nodes);
+            current = RepairRemainingEdgeNodeCrossings(current, nodes);
         }
 
-        // Per-edge gateway scoring opportunity fix: for edges where a shorter
-        // clean exit path is available, apply it directly. Uses the same
-        // lenient validation as the face redirect.
-        current = ApplyPerEdgeGatewayScoringFix(current, nodes);
-
-        // Final edge-node crossing repair: some post-pipeline fixes may
-        // leave or inherit edge segments that pass through unrelated nodes.
-        // Push crossing horizontal segments above/below the blocking node.
-        current = RepairRemainingEdgeNodeCrossings(current, nodes);
-
         return current;
     }
 
@@ -322,25 +315,25 @@ internal static partial class ElkEdgeRouterIterative
             // 19+ boundary-slot violations.
             candidate = ElkEdgePostProcessor.StraightenGatewayCornerDiagonals(candidate, nodes);
 
-            // Check if the fix creates any new hard-rule violations.
+            // Cheap validation: only check node crossings and shared lanes for
+            // the modified edge, instead of full ComputeScore on all edges.
+            // Full scoring per edge is O(edges^2) and causes 2min+ regression.
+            var modifiedEdge = candidate.First(e => string.Equals(e.Id, edgeId, StringComparison.Ordinal));
+            var crossings = ElkEdgeRoutingScoring.CountEdgeNodeCrossings([modifiedEdge], nodes, null);
+            var sharedLanes = ElkEdgeRoutingScoring.CountSharedLaneViolations([modifiedEdge], nodes);
+            if (crossings > 0 || sharedLanes > 0)
+            {
+                continue;
+            }
+
+            // Full score for accepted candidates only (amortized cost).
             var candidateScore = ElkEdgeRoutingScoring.ComputeScore(candidate, nodes);
-            var candidateRetry = BuildRetryState(
+            var candidateRetry = BuildRetryState(candidateScore, 0);
-                candidateScore,
-                HighwayProcessingEnabled
-                    ? ElkEdgeRouterHighway.DetectRemainingBrokenHighways(candidate, nodes).Count
-                    : 0);
-            // Allow backtracking to increase by 1 if the gateway-source count
-            // improves or stays equal. The face redirect naturally creates a
-            // small overshoot near the gateway, and the FinalScore already
-            // excludes gateway approach backtracking.
             var candidateGatewaySourceBetter =
                 candidateRetry.GatewaySourceExitViolations <= current.RetryState.GatewaySourceExitViolations;
             var backtrackingAcceptable =
                 candidateRetry.TargetApproachBacktrackingViolations <= current.RetryState.TargetApproachBacktrackingViolations + 1
                 && candidateGatewaySourceBetter;
-            // Also allow boundary-slots to increase by up to 3 — the redirect
-            // changes the exit point which may temporarily misalign with the
-            // slot lattice. The final boundary-slot snap pass will clean up.
             var boundarySlotAcceptable =
                 candidateRetry.BoundarySlotViolations <= current.RetryState.BoundarySlotViolations + 3
                 && candidateGatewaySourceBetter;
@@ -432,15 +425,18 @@ internal static partial class ElkEdgeRouterIterative
             };
 
             candidateEdges = ElkEdgePostProcessor.StraightenGatewayCornerDiagonals(candidateEdges, nodes);
-            var candidateScore = ElkEdgeRoutingScoring.ComputeScore(candidateEdges, nodes);
-            var candidateRetry = BuildRetryState(
-                candidateScore,
-                HighwayProcessingEnabled
-                    ? ElkEdgeRouterHighway.DetectRemainingBrokenHighways(candidateEdges, nodes).Count
-                    : 0);
 
-            // Lenient: allow backtracking +1 and boundary-slots +3
+            // Cheap validation first: reject if modified edge creates crossings/shared lanes.
-            // (redirect may create small overshoot and shift slot alignment).
+            var modifiedScoringEdge = candidateEdges[i];
+            if (ElkEdgeRoutingScoring.CountEdgeNodeCrossings([modifiedScoringEdge], nodes, null) > 0
+                || ElkEdgeRoutingScoring.CountSharedLaneViolations([modifiedScoringEdge], nodes) > 0)
+            {
+                continue;
+            }
+
+            // Full score only for candidates that pass the cheap check.
+            var candidateScore = ElkEdgeRoutingScoring.ComputeScore(candidateEdges, nodes);
+            var candidateRetry = BuildRetryState(candidateScore, 0);
             var backtrackingOk = candidateRetry.TargetApproachBacktrackingViolations
                 <= current.RetryState.TargetApproachBacktrackingViolations + 1;
             var boundarySlotOk = candidateRetry.BoundarySlotViolations