Optimize WMO rendering and collision performance

- Add texture-sorted batch merging with glMultiDrawElements to reduce draw calls
- Pre-compute merged batches at load time instead of per-frame
- Add persistent floor height cache with disk save/load (cache/wmo_floor_cache.bin)
- Reduce collision focus radius and sweep steps for faster collision checks
- Add floor cache size to performance HUD
- Reduce WMO group distance culling to 80 units

include/rendering/camera_controller.hpp

@@ -106,8 +106,8 @@ private:
     static constexpr float PIVOT_HEIGHT = 1.8f;     // Pivot at head height
     static constexpr float CAM_SPHERE_RADIUS = 0.32f;  // Keep camera farther from geometry to avoid clipping-through surfaces
     static constexpr float CAM_EPSILON = 0.22f;        // Extra wall offset to avoid near-plane clipping artifacts
-    static constexpr float COLLISION_FOCUS_RADIUS_THIRD_PERSON = 42.0f;
-    static constexpr float COLLISION_FOCUS_RADIUS_FREE_FLY = 34.0f;
+    static constexpr float COLLISION_FOCUS_RADIUS_THIRD_PERSON = 20.0f;  // Reduced for performance
+    static constexpr float COLLISION_FOCUS_RADIUS_FREE_FLY = 20.0f;
     static constexpr float MIN_PITCH = -88.0f;      // Look almost straight down
     static constexpr float MAX_PITCH = 35.0f;       // Limited upward look
     glm::vec3* followTarget = nullptr;

include/rendering/wmo_renderer.hpp

@@ -222,6 +222,11 @@ public:
     double getQueryTimeMs() const { return queryTimeMs; }
     uint32_t getQueryCallCount() const { return queryCallCount; }
 
+    // Floor cache persistence
+    bool saveFloorCache(const std::string& filepath) const;
+    bool loadFloorCache(const std::string& filepath);
+    size_t getFloorCacheSize() const { return precomputedFloorGrid.size(); }
+
 private:
     /**
      * WMO group GPU resources
@@ -243,6 +248,16 @@ private:
         };
         std::vector<Batch> batches;
 
+        // Pre-merged batches for efficient rendering (computed at load time)
+        struct MergedBatch {
+            GLuint texId;
+            bool hasTexture;
+            bool alphaTest;
+            std::vector<GLsizei> counts;
+            std::vector<const void*> offsets;
+        };
+        std::vector<MergedBatch> mergedBatches;
+
         // Collision geometry (positions only, for floor raycasting)
         std::vector<glm::vec3> collisionVertices;
         std::vector<uint16_t> collisionIndices;
@@ -472,6 +487,22 @@ private:
     // Collision query profiling (per frame).
     mutable double queryTimeMs = 0.0;
     mutable uint32_t queryCallCount = 0;
+
+    // Floor height cache - persistent precomputed grid
+    static constexpr float FLOOR_GRID_CELL_SIZE = 2.0f;  // 2 unit grid cells
+    mutable std::unordered_map<uint64_t, float> precomputedFloorGrid;  // key -> floor height
+    mutable bool floorGridDirty = true;  // Rebuild when instances change
+    mutable uint32_t currentFrameId = 0;
+
+    uint64_t floorGridKey(float x, float y) const {
+        int32_t ix = static_cast<int32_t>(std::floor(x / FLOOR_GRID_CELL_SIZE));
+        int32_t iy = static_cast<int32_t>(std::floor(y / FLOOR_GRID_CELL_SIZE));
+        return (static_cast<uint64_t>(static_cast<uint32_t>(ix)) << 32) |
+               static_cast<uint64_t>(static_cast<uint32_t>(iy));
+    }
+
+    // Compute floor height for a single cell (expensive, done at load time)
+    std::optional<float> computeFloorHeightSlow(float x, float y, float refZ) const;
 };
 
 } // namespace rendering