Overhaul WMO collision: precompute normals, fix floor selection, optimize queries

- Precompute triangle normals in buildCollisionGrid, eliminating per-query
  cross+normalize in getFloorHeight, checkWallCollision, and raycastBoundingBoxes
- Fix floor selection: remove redundant allowAbove (callers already elevate
  probeZ by stepUpBudget), preventing upper-story snap at doorway transitions
- Align wall classification threshold (absNz < 0.35) with runtime skip check,
  eliminating ~30% wasted wall triangle fetches
- Replace O(n log n) sort+unique dedup in range queries with O(n) visited bitset
- Rename wallTriScratch to triScratch_, fix stale threshold comments

include/rendering/wmo_renderer.hpp

@@ -390,13 +390,19 @@ private:
         std::vector<std::vector<uint32_t>> cellTriangles;
 
         // Pre-classified triangle lists per cell (built at load time)
-        std::vector<std::vector<uint32_t>> cellFloorTriangles;  // abs(normal.z) >= 0.45
-        std::vector<std::vector<uint32_t>> cellWallTriangles;   // abs(normal.z) < 0.55
+        std::vector<std::vector<uint32_t>> cellFloorTriangles;  // abs(normal.z) >= 0.35
+        std::vector<std::vector<uint32_t>> cellWallTriangles;   // abs(normal.z) < 0.35
 
         // Pre-computed per-triangle Z bounds for fast vertical reject
         struct TriBounds { float minZ; float maxZ; };
         std::vector<TriBounds> triBounds;  // indexed by triStart/3
 
+        // Pre-computed per-triangle normals (unit length, indexed by triStart/3)
+        std::vector<glm::vec3> triNormals;
+
+        // Scratch bitset for deduplicating triangle queries (sized to numTriangles)
+        mutable std::vector<uint8_t> triVisited;
+
         // Build the spatial grid from collision geometry
         void buildCollisionGrid();
 
@@ -675,7 +681,7 @@ private:
     std::unordered_map<GridCell, std::vector<uint32_t>, GridCellHash> spatialGrid;
     std::unordered_map<uint32_t, size_t> instanceIndexById;
     mutable std::vector<size_t> candidateScratch;
-    mutable std::vector<uint32_t> wallTriScratch;  // Scratch for wall collision grid queries
+    mutable std::vector<uint32_t> triScratch_;  // Scratch for collision grid queries
     mutable std::unordered_set<uint32_t> candidateIdScratch;
 
     // Parallel visibility culling