Fix WMO wall collision, normal mapping, POM backfill, and M2/WMO rendering performance

- Fix MOPY flag check (0x08 not 0x01) for proper wall collision detection
- Cap MAX_PUSH to PLAYER_RADIUS to prevent gradual clip-through
- Fix WMO doodad quaternion component ordering (X/Y swap)
- Linear normal map strength blend in shader for smooth slider control
- Enable shadow sampling for interior WMO groups (covered outdoor areas)
- Backfill deferred normal/height maps after streaming with descriptor rebind
- M2: prepareRender only iterates animated instances, bone dirty flag
- M2: remove worker thread VMA allocation, skip unready bone instances
- WMO: persistent visibility vectors, sequential culling
- Add FSR EASU/RCAS shaders

assets/shaders/fsr_easu.frag.glsl

@@ -0,0 +1,102 @@
+#version 450
+// FSR 1.0 EASU (Edge Adaptive Spatial Upsampling) — Fragment Shader
+// Based on AMD FidelityFX Super Resolution 1.0
+// Implements edge-adaptive bilinear upsampling with directional filtering
+
+layout(set = 0, binding = 0) uniform sampler2D uInput;
+
+layout(push_constant) uniform FSRConstants {
+    vec4 con0; // inputSize.xy, 1/inputSize.xy
+    vec4 con1; // inputSize.xy / outputSize.xy, 0.5 * inputSize.xy / outputSize.xy
+    vec4 con2; // outputSize.xy, 1/outputSize.xy
+    vec4 con3; // sharpness, 0, 0, 0
+} fsr;
+
+layout(location = 0) in vec2 TexCoord;
+layout(location = 0) out vec4 outColor;
+
+// Fetch a texel with offset (in input pixels)
+vec3 fsrFetch(vec2 p, vec2 off) {
+    return textureLod(uInput, (p + off + 0.5) * fsr.con0.zw, 0.0).rgb;
+}
+
+void main() {
+    // Undo the vertex shader Y flip (postprocess.vert flips for Vulkan overlay,
+    // but we need standard UV coords for texture sampling)
+    vec2 tc = vec2(TexCoord.x, 1.0 - TexCoord.y);
+
+    // Map output pixel to input space
+    vec2 pp = tc * fsr.con2.xy; // output pixel position
+    vec2 ip = pp * fsr.con1.xy - 0.5; // input pixel position (centered)
+    vec2 fp = floor(ip);
+    vec2 ff = ip - fp;
+
+    // 12-tap filter: 4x3 grid around the pixel
+    //  b c
+    // e f g h
+    // i j k l
+    //  n o
+    vec3 b = fsrFetch(fp, vec2( 0, -1));
+    vec3 c = fsrFetch(fp, vec2( 1, -1));
+    vec3 e = fsrFetch(fp, vec2(-1,  0));
+    vec3 f = fsrFetch(fp, vec2( 0,  0));
+    vec3 g = fsrFetch(fp, vec2( 1,  0));
+    vec3 h = fsrFetch(fp, vec2( 2,  0));
+    vec3 i = fsrFetch(fp, vec2(-1,  1));
+    vec3 j = fsrFetch(fp, vec2( 0,  1));
+    vec3 k = fsrFetch(fp, vec2( 1,  1));
+    vec3 l = fsrFetch(fp, vec2( 2,  1));
+    vec3 n = fsrFetch(fp, vec2( 0,  2));
+    vec3 o = fsrFetch(fp, vec2( 1,  2));
+
+    // Luma (use green channel as good perceptual approximation)
+    float bL = b.g, cL = c.g, eL = e.g, fL = f.g;
+    float gL = g.g, hL = h.g, iL = i.g, jL = j.g;
+    float kL = k.g, lL = l.g, nL = n.g, oL = o.g;
+
+    // Directional edge detection
+    // Compute gradients in 4 directions (N-S, E-W, NE-SW, NW-SE)
+    float dc = cL - jL;
+    float db = bL - kL;
+    float de = eL - hL;
+    float di = iL - lL;
+
+    // Length of the edge in each direction
+    float lenH = abs(eL - fL) + abs(fL - gL) + abs(iL - jL) + abs(jL - kL);
+    float lenV = abs(bL - fL) + abs(fL - jL) + abs(cL - gL) + abs(gL - kL);
+
+    // Determine dominant edge direction
+    float dirH = lenV / (lenH + lenV + 1e-7);
+    float dirV = lenH / (lenH + lenV + 1e-7);
+
+    // Bilinear weights
+    float w1 = (1.0 - ff.x) * (1.0 - ff.y);
+    float w2 = ff.x * (1.0 - ff.y);
+    float w3 = (1.0 - ff.x) * ff.y;
+    float w4 = ff.x * ff.y;
+
+    // Edge-aware sharpening: boost weights along edges
+    float sharpness = fsr.con3.x;
+    float edgeStr = max(abs(lenH - lenV) / (lenH + lenV + 1e-7), 0.0);
+    float sharp = mix(0.0, sharpness, edgeStr);
+
+    // Sharpen bilinear by pulling toward nearest texel
+    float maxW = max(max(w1, w2), max(w3, w4));
+    w1 = mix(w1, float(w1 == maxW), sharp * 0.25);
+    w2 = mix(w2, float(w2 == maxW), sharp * 0.25);
+    w3 = mix(w3, float(w3 == maxW), sharp * 0.25);
+    w4 = mix(w4, float(w4 == maxW), sharp * 0.25);
+
+    // Normalize
+    float wSum = w1 + w2 + w3 + w4;
+    w1 /= wSum; w2 /= wSum; w3 /= wSum; w4 /= wSum;
+
+    // Final color: weighted blend of the 4 nearest texels with edge awareness
+    vec3 color = f * w1 + g * w2 + j * w3 + k * w4;
+
+    // Optional: blend in some of the surrounding texels for anti-aliasing
+    float aa = 0.125 * edgeStr;
+    color = mix(color, (b + c + e + h + i + l + n + o) / 8.0, aa * 0.15);
+
+    outColor = vec4(clamp(color, 0.0, 1.0), 1.0);
+}

assets/shaders/fsr_rcas.frag.glsl

@@ -0,0 +1,43 @@
+#version 450
+// FSR 1.0 RCAS (Robust Contrast Adaptive Sharpening) — Fragment Shader
+// Based on AMD FidelityFX Super Resolution 1.0
+// Applies contrast-adaptive sharpening after EASU upscaling
+
+layout(set = 0, binding = 0) uniform sampler2D uInput;
+
+layout(push_constant) uniform RCASConstants {
+    vec4 con0; // 1/outputSize.xy, outputSize.xy
+    vec4 con1; // sharpness (x), 0, 0, 0
+} rcas;
+
+layout(location = 0) in vec2 TexCoord;
+layout(location = 0) out vec4 outColor;
+
+void main() {
+    // Fetch center and 4-neighborhood
+    vec2 texelSize = rcas.con0.xy;
+    vec3 c = texture(uInput, TexCoord).rgb;
+    vec3 n = texture(uInput, TexCoord + vec2( 0, -texelSize.y)).rgb;
+    vec3 s = texture(uInput, TexCoord + vec2( 0,  texelSize.y)).rgb;
+    vec3 w = texture(uInput, TexCoord + vec2(-texelSize.x,  0)).rgb;
+    vec3 e = texture(uInput, TexCoord + vec2( texelSize.x,  0)).rgb;
+
+    // Luma (green channel approximation)
+    float cL = c.g, nL = n.g, sL = s.g, wL = w.g, eL = e.g;
+
+    // Min/max of neighborhood
+    float minL = min(min(nL, sL), min(wL, eL));
+    float maxL = max(max(nL, sL), max(wL, eL));
+
+    // Contrast adaptive sharpening weight
+    // Higher contrast = less sharpening to avoid ringing
+    float contrast = maxL - minL;
+    float sharpness = rcas.con1.x;
+    float w0 = sharpness * (1.0 - smoothstep(0.0, 0.3, contrast));
+
+    // Apply sharpening: center + w0 * (center - average_neighbors)
+    vec3 avg = (n + s + w + e) * 0.25;
+    vec3 sharpened = c + w0 * (c - avg);
+
+    outColor = vec4(clamp(sharpened, 0.0, 1.0), 1.0);
+}

assets/shaders/wmo.frag.glsl

@@ -149,21 +149,21 @@ void main() {
     vec3 norm = vertexNormal;
     if (enableNormalMap != 0 && lodFactor < 0.99 && normalMapStrength > 0.001) {
         vec3 mapNormal = texture(uNormalHeightMap, finalUV).rgb * 2.0 - 1.0;
-        // Scale XY by strength to control effect intensity
-        mapNormal.xy *= normalMapStrength;
         mapNormal = normalize(mapNormal);
         vec3 worldNormal = normalize(TBN * mapNormal);
         if (!gl_FrontFacing) worldNormal = -worldNormal;
-        // Blend: strength + LOD both contribute to fade toward vertex normal
-        float blendFactor = max(lodFactor, 1.0 - normalMapStrength);
-        norm = normalize(mix(worldNormal, vertexNormal, blendFactor));
+        // Linear blend: strength controls how much normal map detail shows,
+        // LOD fades out at distance. Both multiply for smooth falloff.
+        float blend = clamp(normalMapStrength, 0.0, 1.0) * (1.0 - lodFactor);
+        norm = normalize(mix(vertexNormal, worldNormal, blend));
     }
 
     vec3 result;
 
-    // Sample shadow map — skip for interior WMO groups (no sun indoors)
+    // Sample shadow map for all WMO groups (interior groups with 0x2000 flag
+    // include covered outdoor areas like archways/streets that should receive shadows)
     float shadow = 1.0;
-    if (shadowParams.x > 0.5 && isInterior == 0) {
+    if (shadowParams.x > 0.5) {
         vec3 ldir = normalize(-lightDir.xyz);
         float normalOffset = SHADOW_TEXEL * 2.0 * (1.0 - abs(dot(norm, ldir)));
         vec3 biasedPos = FragPos + norm * normalOffset;

include/core/application.hpp

@@ -236,6 +236,11 @@ private:
     std::optional<PendingWorldEntry> pendingWorldEntry_;  // Deferred world entry during loading
     float taxiLandingClampTimer_ = 0.0f;
     float worldEntryMovementGraceTimer_ = 0.0f;
+
+    // Hearth teleport: freeze player until terrain loads at destination
+    bool hearthTeleportPending_ = false;
+    glm::vec3 hearthTeleportPos_{0.0f};  // render coords
+    float hearthTeleportTimer_ = 0.0f;   // timeout safety
     float facingSendCooldown_ = 0.0f;        // Rate-limits MSG_MOVE_SET_FACING
     float lastSentCanonicalYaw_ = 1000.0f;   // Sentinel — triggers first send
     float taxiStreamCooldown_ = 0.0f;

include/game/game_handler.hpp

@@ -565,6 +565,8 @@ public:
     void unstuck();
     void setUnstuckGyCallback(UnstuckCallback cb) { unstuckGyCallback_ = std::move(cb); }
     void unstuckGy();
+    void setUnstuckHearthCallback(UnstuckCallback cb) { unstuckHearthCallback_ = std::move(cb); }
+    void unstuckHearth();
     using BindPointCallback = std::function<void(uint32_t mapId, float x, float y, float z)>;
     void setBindPointCallback(BindPointCallback cb) { bindPointCallback_ = std::move(cb); }
 
@@ -1445,6 +1447,7 @@ private:
     WorldEntryCallback worldEntryCallback_;
     UnstuckCallback unstuckCallback_;
     UnstuckCallback unstuckGyCallback_;
+    UnstuckCallback unstuckHearthCallback_;
     BindPointCallback bindPointCallback_;
     CreatureSpawnCallback creatureSpawnCallback_;
     CreatureDespawnCallback creatureDespawnCallback_;

include/rendering/character_renderer.hpp

@@ -66,6 +66,8 @@ public:
 
     void update(float deltaTime, const glm::vec3& cameraPos = glm::vec3(0.0f));
 
+    /** Pre-allocate GPU resources (bone SSBOs, descriptors) on main thread before parallel render. */
+    void prepareRender(uint32_t frameIndex);
     void render(VkCommandBuffer cmd, VkDescriptorSet perFrameSet, const Camera& camera);
     void recreatePipelines();
     bool initializeShadow(VkRenderPass shadowRenderPass);

include/rendering/m2_renderer.hpp

@@ -122,6 +122,7 @@ struct M2ModelGPU {
     bool isKoboldFlame = false;     // Model name matches kobold+(candle/torch/mine) (precomputed)
     bool isLavaModel = false;       // Model name contains lava/molten/magma (UV scroll fallback)
     bool hasTextureAnimation = false; // True if any batch has UV animation
+    uint8_t availableLODs = 0;  // Bitmask: bit N set if any batch has submeshLevel==N
 
     // Particle emitter data (kept from M2Model)
     std::vector<pipeline::M2ParticleEmitter> particleEmitters;
@@ -193,6 +194,7 @@ struct M2Instance {
 
     // Frame-skip optimization (update distant animations less frequently)
     uint8_t frameSkipCounter = 0;
+    bool bonesDirty = false;  // Set when bones recomputed, cleared after upload
 
     // Per-instance bone SSBO (double-buffered)
     ::VkBuffer boneBuffer[2] = {};
@@ -265,6 +267,8 @@ public:
     /**
      * Render all visible instances (Vulkan)
      */
+    /** Pre-allocate GPU resources (bone SSBOs, descriptors) on main thread before parallel render. */
+    void prepareRender(uint32_t frameIndex, const Camera& camera);
     void render(VkCommandBuffer cmd, VkDescriptorSet perFrameSet, const Camera& camera);
 
     /**

include/rendering/renderer.hpp

@@ -4,10 +4,12 @@
 #include <string>
 #include <cstdint>
 #include <vector>
+#include <future>
 #include <glm/glm.hpp>
 #include <vulkan/vulkan.h>
 #include <vk_mem_alloc.h>
 #include "rendering/vk_frame_data.hpp"
+#include "rendering/vk_utils.hpp"
 #include "rendering/sky_system.hpp"
 
 namespace wowee {
@@ -259,6 +261,14 @@ public:
     float getShadowDistance() const { return shadowDistance_; }
     void setMsaaSamples(VkSampleCountFlagBits samples);
 
+    // FSR 1.0 (FidelityFX Super Resolution) upscaling
+    void setFSREnabled(bool enabled);
+    bool isFSREnabled() const { return fsr_.enabled; }
+    void setFSRQuality(float scaleFactor);  // 0.50=Perf, 0.59=Balanced, 0.67=Quality, 0.77=UltraQuality
+    void setFSRSharpness(float sharpness);  // 0.0 - 2.0
+    float getFSRScaleFactor() const { return fsr_.scaleFactor; }
+    float getFSRSharpness() const { return fsr_.sharpness; }
+
     void setWaterRefractionEnabled(bool enabled);
     bool isWaterRefractionEnabled() const;
 
@@ -312,7 +322,7 @@ private:
     VmaAllocation selCircleIdxAlloc = VK_NULL_HANDLE;
     int selCircleVertCount = 0;
     void initSelectionCircle();
-    void renderSelectionCircle(const glm::mat4& view, const glm::mat4& projection);
+    void renderSelectionCircle(const glm::mat4& view, const glm::mat4& projection, VkCommandBuffer overrideCmd = VK_NULL_HANDLE);
     glm::vec3 selCirclePos{0.0f};
     glm::vec3 selCircleColor{1.0f, 0.0f, 0.0f};
     float selCircleRadius = 1.5f;
@@ -322,7 +332,36 @@ private:
     VkPipeline overlayPipeline = VK_NULL_HANDLE;
     VkPipelineLayout overlayPipelineLayout = VK_NULL_HANDLE;
     void initOverlayPipeline();
-    void renderOverlay(const glm::vec4& color);
+    void renderOverlay(const glm::vec4& color, VkCommandBuffer overrideCmd = VK_NULL_HANDLE);
+
+    // FSR 1.0 upscaling state
+    struct FSRState {
+        bool enabled = false;
+        bool needsRecreate = false;
+        float scaleFactor = 0.77f;  // Ultra Quality default
+        float sharpness = 0.5f;
+        uint32_t internalWidth = 0;
+        uint32_t internalHeight = 0;
+
+        // Off-screen scene target (reduced resolution)
+        AllocatedImage sceneColor{};        // 1x color (non-MSAA render target / MSAA resolve target)
+        AllocatedImage sceneDepth{};        // Depth (matches current MSAA sample count)
+        AllocatedImage sceneMsaaColor{};    // MSAA color target (only when MSAA > 1x)
+        AllocatedImage sceneDepthResolve{}; // Depth resolve (only when MSAA + depth resolve)
+        VkFramebuffer sceneFramebuffer = VK_NULL_HANDLE;
+        VkSampler sceneSampler = VK_NULL_HANDLE;
+
+        // Upscale pipeline
+        VkPipeline pipeline = VK_NULL_HANDLE;
+        VkPipelineLayout pipelineLayout = VK_NULL_HANDLE;
+        VkDescriptorSetLayout descSetLayout = VK_NULL_HANDLE;
+        VkDescriptorPool descPool = VK_NULL_HANDLE;
+        VkDescriptorSet descSet = VK_NULL_HANDLE;
+    };
+    FSRState fsr_;
+    bool initFSRResources();
+    void destroyFSRResources();
+    void renderFSRUpscale();
 
     // Footstep event tracking (animation-driven)
     uint32_t footstepLastAnimationId = 0;
@@ -411,6 +450,36 @@ private:
     void setupWater1xPass();
     void renderReflectionPass();
 
+    // ── Multithreaded secondary command buffer recording ──
+    // Indices into secondaryCmds_ arrays
+    static constexpr uint32_t SEC_SKY     = 0;  // sky (main thread)
+    static constexpr uint32_t SEC_TERRAIN = 1;  // terrain (worker 0)
+    static constexpr uint32_t SEC_WMO     = 2;  // WMO (worker 1)
+    static constexpr uint32_t SEC_CHARS   = 3;  // selection circle + characters (main thread)
+    static constexpr uint32_t SEC_M2      = 4;  // M2 + particles + glow (worker 2)
+    static constexpr uint32_t SEC_POST    = 5;  // water + weather + effects (main thread)
+    static constexpr uint32_t SEC_IMGUI   = 6;  // ImGui (main thread, non-FSR only)
+    static constexpr uint32_t NUM_SECONDARIES = 7;
+    static constexpr uint32_t NUM_WORKERS = 3;  // terrain, WMO, M2
+
+    // Per-worker command pools (thread-safe: one pool per thread)
+    VkCommandPool workerCmdPools_[NUM_WORKERS] = {};
+    // Main-thread command pool for its secondary buffers
+    VkCommandPool mainSecondaryCmdPool_ = VK_NULL_HANDLE;
+    // Pre-allocated secondary command buffers [secondaryIndex][frameInFlight]
+    VkCommandBuffer secondaryCmds_[NUM_SECONDARIES][MAX_FRAMES] = {};
+
+    bool parallelRecordingEnabled_ = false;  // set true after pools/buffers created
+    bool createSecondaryCommandResources();
+    void destroySecondaryCommandResources();
+    VkCommandBuffer beginSecondary(uint32_t secondaryIndex);
+    void setSecondaryViewportScissor(VkCommandBuffer cmd);
+
+    // Cached render pass state for secondary buffer inheritance
+    VkRenderPass activeRenderPass_ = VK_NULL_HANDLE;
+    VkFramebuffer activeFramebuffer_ = VK_NULL_HANDLE;
+    VkExtent2D activeRenderExtent_ = {0, 0};
+
     // Active character previews for off-screen rendering
     std::vector<CharacterPreview*> activePreviews_;
 

include/rendering/vk_context.hpp

@@ -84,6 +84,10 @@ public:
     bool isSwapchainDirty() const { return swapchainDirty; }
     void markSwapchainDirty() { swapchainDirty = true; }
 
+    // VSync (present mode)
+    bool isVsyncEnabled() const { return vsync_; }
+    void setVsync(bool enabled) { vsync_ = enabled; }
+
     bool isDeviceLost() const { return deviceLost_; }
 
     // MSAA
@@ -145,6 +149,7 @@ private:
     std::vector<VkFramebuffer> swapchainFramebuffers;
     bool swapchainDirty = false;
     bool deviceLost_ = false;
+    bool vsync_ = true;
 
     // Per-frame resources
     FrameData frames[MAX_FRAMES_IN_FLIGHT];

include/rendering/wmo_renderer.hpp

@@ -148,6 +148,8 @@ public:
      * @param perFrameSet Per-frame descriptor set (set 0)
      * @param camera Camera for frustum culling
      */
+    /** Pre-update mutable state (frame ID, material UBOs) on main thread before parallel render. */
+    void prepareRender();
     void render(VkCommandBuffer cmd, VkDescriptorSet perFrameSet, const Camera& camera);
 
     /**
@@ -332,6 +334,9 @@ public:
     // Defer normal/height map generation during streaming to avoid CPU stalls
     void setDeferNormalMaps(bool defer) { deferNormalMaps_ = defer; }
 
+    // Generate normal/height maps for cached textures that were loaded while deferred
+    void backfillNormalMaps();
+
 private:
     // WMO material UBO — matches WMOMaterial in wmo.frag.glsl
     struct WMOMaterialUBO {
@@ -720,6 +725,8 @@ private:
         uint32_t distanceCulled = 0;
     };
     std::vector<std::future<void>> cullFutures_;
+    std::vector<size_t> visibleInstances_;      // reused per frame
+    std::vector<InstanceDrawList> drawLists_;    // reused per frame
 
     // Collision query profiling (per frame).
     mutable double queryTimeMs = 0.0;

include/ui/game_screen.hpp

@@ -116,6 +116,10 @@ private:
     float pendingNormalMapStrength = 0.8f;  // 0.0-2.0
     bool pendingPOM = true;             // on by default
     int pendingPOMQuality = 1;          // 0=Low(16), 1=Medium(32), 2=High(64)
+    bool pendingFSR = false;
+    int pendingFSRQuality = 0;          // 0=UltraQuality, 1=Quality, 2=Balanced, 3=Performance
+    float pendingFSRSharpness = 0.5f;
+    bool fsrSettingsApplied_ = false;
 
     // UI element transparency (0.0 = fully transparent, 1.0 = fully opaque)
     float uiOpacity_ = 0.65f;

src/core/application.cpp

@@ -1015,14 +1015,33 @@ void Application::update(float deltaTime) {
                     if (renderer && renderer->getCameraController())
                         renderer->getCameraController()->clearMovementInputs();
                 }
+                // Hearth teleport: keep player frozen until terrain loads at destination
+                if (hearthTeleportPending_ && renderer && renderer->getTerrainManager()) {
+                    hearthTeleportTimer_ -= deltaTime;
+                    auto terrainH = renderer->getTerrainManager()->getHeightAt(
+                        hearthTeleportPos_.x, hearthTeleportPos_.y);
+                    if (terrainH || hearthTeleportTimer_ <= 0.0f) {
+                        // Terrain loaded (or timeout) — snap to floor and release
+                        if (terrainH) {
+                            hearthTeleportPos_.z = *terrainH + 0.5f;
+                            renderer->getCameraController()->teleportTo(hearthTeleportPos_);
+                        }
+                        renderer->getCameraController()->setExternalFollow(false);
+                        worldEntryMovementGraceTimer_ = 1.0f;
+                        hearthTeleportPending_ = false;
+                        LOG_INFO("Unstuck hearth: terrain loaded, player released",
+                                 terrainH ? "" : " (timeout)");
+                    }
+                }
                 if (renderer && renderer->getCameraController()) {
                 const bool externallyDrivenMotion = onTaxi || onWMOTransport || chargeActive_;
                 // Keep physics frozen (externalFollow) during landing clamp when terrain
                 // hasn't loaded yet — prevents gravity from pulling player through void.
+                bool hearthFreeze = hearthTeleportPending_;
                 bool landingClampActive = !onTaxi && taxiLandingClampTimer_ > 0.0f &&
                                           worldEntryMovementGraceTimer_ <= 0.0f &&
                                           !gameHandler->isMounted();
-                renderer->getCameraController()->setExternalFollow(externallyDrivenMotion || landingClampActive);
+                renderer->getCameraController()->setExternalFollow(externallyDrivenMotion || landingClampActive || hearthFreeze);
                 renderer->getCameraController()->setExternalMoving(externallyDrivenMotion);
                 if (externallyDrivenMotion) {
                     // Drop any stale local movement toggles while server drives taxi motion.
@@ -1877,9 +1896,43 @@ void Application::setupUICallbacks() {
         LOG_INFO("Unstuck: high fallback snap");
     });
 
+    // /unstuckhearth — teleport to hearthstone bind point (server-synced).
+    // Freezes player until terrain loads at destination to prevent falling through world.
+    gameHandler->setUnstuckHearthCallback([this, clearStuckMovement, forceServerTeleportCommand]() {
+        if (!renderer || !renderer->getCameraController() || !gameHandler) return;
+
+        uint32_t bindMap = 0;
+        glm::vec3 bindPos(0.0f);
+        if (!gameHandler->getHomeBind(bindMap, bindPos)) {
+            LOG_WARNING("Unstuck hearth: no bind point available");
+            return;
+        }
+
+        worldEntryMovementGraceTimer_ = 10.0f;  // long grace — terrain load check will clear it
+        taxiLandingClampTimer_ = 0.0f;
+        lastTaxiFlight_ = false;
+        clearStuckMovement();
+
+        auto* cc = renderer->getCameraController();
+        glm::vec3 renderPos = core::coords::canonicalToRender(bindPos);
+        renderPos.z += 2.0f;
+
+        // Freeze player in place (no gravity/movement) until terrain loads
+        cc->teleportTo(renderPos);
+        cc->setExternalFollow(true);
+        forceServerTeleportCommand(renderPos);
+        clearStuckMovement();
+
+        // Set pending state — update loop will unfreeze once terrain is loaded
+        hearthTeleportPending_ = true;
+        hearthTeleportPos_ = renderPos;
+        hearthTeleportTimer_ = 15.0f;  // 15s safety timeout
+        LOG_INFO("Unstuck hearth: teleporting to bind point, waiting for terrain...");
+    });
+
     // Auto-unstuck: falling for > 5 seconds = void fall, teleport to map entry
     if (renderer->getCameraController()) {
-        renderer->getCameraController()->setAutoUnstuckCallback([this]() {
+        renderer->getCameraController()->setAutoUnstuckCallback([this, forceServerTeleportCommand]() {
             if (!renderer || !renderer->getCameraController()) return;
             auto* cc = renderer->getCameraController();
 
@@ -1887,7 +1940,8 @@ void Application::setupUICallbacks() {
             glm::vec3 spawnPos = cc->getDefaultPosition();
             spawnPos.z += 5.0f;
             cc->teleportTo(spawnPos);
-            LOG_INFO("Auto-unstuck: teleported to map entry point");
+            forceServerTeleportCommand(spawnPos);
+            LOG_INFO("Auto-unstuck: teleported to map entry point (server synced)");
         });
     }
 

src/core/window.cpp

@@ -84,6 +84,7 @@ bool Window::initialize() {
 
     // Initialize Vulkan context
     vkContext = std::make_unique<rendering::VkContext>();
+    vkContext->setVsync(vsync);
     if (!vkContext->initialize(window)) {
         LOG_ERROR("Failed to initialize Vulkan context");
         return false;
@@ -158,11 +159,13 @@ void Window::setFullscreen(bool enable) {
     }
 }
 
-void Window::setVsync([[maybe_unused]] bool enable) {
-    // VSync in Vulkan is controlled by present mode (set at swapchain creation)
-    // For now, store the preference — applied on next swapchain recreation
+void Window::setVsync(bool enable) {
     vsync = enable;
-    LOG_INFO("VSync preference set to ", enable ? "on" : "off", " (applied on swapchain recreation)");
+    if (vkContext) {
+        vkContext->setVsync(enable);
+        vkContext->markSwapchainDirty();
+    }
+    LOG_INFO("VSync ", enable ? "enabled" : "disabled");
 }
 
 void Window::applyResolution(int w, int h) {

src/game/game_handler.cpp

@@ -11435,6 +11435,15 @@ void GameHandler::unstuckGy() {
     }
 }
 
+void GameHandler::unstuckHearth() {
+    if (unstuckHearthCallback_) {
+        unstuckHearthCallback_();
+        addSystemChatMessage("Unstuck: teleported to hearthstone location.");
+    } else {
+        addSystemChatMessage("No hearthstone bind point set.");
+    }
+}
+
 void GameHandler::handleLootResponse(network::Packet& packet) {
     if (!LootResponseParser::parse(packet, currentLoot)) return;
     lootWindowOpen = true;

src/rendering/character_renderer.cpp

@@ -1924,6 +1924,61 @@ glm::mat4 CharacterRenderer::getBoneTransform(const pipeline::M2Bone& bone, floa
 
 // --- Rendering ---
 
+void CharacterRenderer::prepareRender(uint32_t frameIndex) {
+    if (instances.empty() || !opaquePipeline_) return;
+
+    // Pre-allocate bone SSBOs + descriptor sets on main thread (pool ops not thread-safe)
+    for (auto& [id, instance] : instances) {
+        int numBones = std::min(static_cast<int>(instance.boneMatrices.size()), MAX_BONES);
+        if (numBones <= 0) continue;
+
+        if (!instance.boneBuffer[frameIndex]) {
+            VkBufferCreateInfo bci{VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO};
+            bci.size = MAX_BONES * sizeof(glm::mat4);
+            bci.usage = VK_BUFFER_USAGE_STORAGE_BUFFER_BIT;
+            VmaAllocationCreateInfo aci{};
+            aci.usage = VMA_MEMORY_USAGE_CPU_TO_GPU;
+            aci.flags = VMA_ALLOCATION_CREATE_MAPPED_BIT;
+            VmaAllocationInfo allocInfo{};
+            vmaCreateBuffer(vkCtx_->getAllocator(), &bci, &aci,
+                            &instance.boneBuffer[frameIndex], &instance.boneAlloc[frameIndex], &allocInfo);
+            instance.boneMapped[frameIndex] = allocInfo.pMappedData;
+
+            VkDescriptorSetAllocateInfo ai{VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO};
+            ai.descriptorPool = boneDescPool_;
+            ai.descriptorSetCount = 1;
+            ai.pSetLayouts = &boneSetLayout_;
+            VkResult dsRes = vkAllocateDescriptorSets(vkCtx_->getDevice(), &ai, &instance.boneSet[frameIndex]);
+            if (dsRes != VK_SUCCESS) {
+                LOG_ERROR("CharacterRenderer::prepareRender: bone descriptor alloc failed (instance=",
+                          id, ", frame=", frameIndex, ", vk=", static_cast<int>(dsRes), ")");
+                if (instance.boneBuffer[frameIndex]) {
+                    vmaDestroyBuffer(vkCtx_->getAllocator(),
+                                     instance.boneBuffer[frameIndex], instance.boneAlloc[frameIndex]);
+                    instance.boneBuffer[frameIndex] = VK_NULL_HANDLE;
+                    instance.boneAlloc[frameIndex] = VK_NULL_HANDLE;
+                    instance.boneMapped[frameIndex] = nullptr;
+                }
+                continue;
+            }
+
+            if (instance.boneSet[frameIndex]) {
+                VkDescriptorBufferInfo bufInfo{};
+                bufInfo.buffer = instance.boneBuffer[frameIndex];
+                bufInfo.offset = 0;
+                bufInfo.range = bci.size;
+                VkWriteDescriptorSet write{VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET};
+                write.dstSet = instance.boneSet[frameIndex];
+                write.dstBinding = 0;
+                write.descriptorCount = 1;
+                write.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
+                write.pBufferInfo = &bufInfo;
+                vkUpdateDescriptorSets(vkCtx_->getDevice(), 1, &write, 0, nullptr);
+            }
+        }
+    }
+}
+
 void CharacterRenderer::render(VkCommandBuffer cmd, VkDescriptorSet perFrameSet, [[maybe_unused]] const Camera& camera) {
     if (instances.empty() || !opaquePipeline_) {
         return;

src/rendering/m2_renderer.cpp

@@ -1602,6 +1602,12 @@ bool M2Renderer::loadModel(const pipeline::M2Model& model, uint32_t modelId) {
         }
     }
 
+    // Pre-compute available LOD levels to avoid per-instance batch iteration
+    gpuModel.availableLODs = 0;
+    for (const auto& b : gpuModel.batches) {
+        if (b.submeshLevel < 8) gpuModel.availableLODs |= (1u << b.submeshLevel);
+    }
+
     models[modelId] = std::move(gpuModel);
 
     LOG_DEBUG("Loaded M2 model: ", model.name, " (", models[modelId].vertexCount, " vertices, ",
@@ -1911,6 +1917,7 @@ static void computeBoneMatrices(const M2ModelGPU& model, M2Instance& instance) {
             instance.boneMatrices[i] = local;
         }
     }
+    instance.bonesDirty = true;
 }
 
 void M2Renderer::update(float deltaTime, const glm::vec3& cameraPos, const glm::mat4& viewProjection) {
@@ -2172,6 +2179,48 @@ void M2Renderer::update(float deltaTime, const glm::vec3& cameraPos, const glm::
 
 }
 
+void M2Renderer::prepareRender(uint32_t frameIndex, const Camera& camera) {
+    if (!initialized_ || instances.empty()) return;
+    (void)camera;  // reserved for future frustum-based culling
+
+    // Pre-allocate bone SSBOs + descriptor sets on main thread (pool ops not thread-safe).
+    // Only iterate animated instances — static doodads don't need bone buffers.
+    for (size_t idx : animatedInstanceIndices_) {
+        if (idx >= instances.size()) continue;
+        auto& instance = instances[idx];
+
+        if (instance.boneMatrices.empty()) continue;
+
+        if (!instance.boneBuffer[frameIndex]) {
+            VkBufferCreateInfo bci{VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO};
+            bci.size = 128 * sizeof(glm::mat4);
+            bci.usage = VK_BUFFER_USAGE_STORAGE_BUFFER_BIT;
+            VmaAllocationCreateInfo aci{};
+            aci.usage = VMA_MEMORY_USAGE_CPU_TO_GPU;
+            aci.flags = VMA_ALLOCATION_CREATE_MAPPED_BIT;
+            VmaAllocationInfo allocInfo{};
+            vmaCreateBuffer(vkCtx_->getAllocator(), &bci, &aci,
+                            &instance.boneBuffer[frameIndex], &instance.boneAlloc[frameIndex], &allocInfo);
+            instance.boneMapped[frameIndex] = allocInfo.pMappedData;
+
+            instance.boneSet[frameIndex] = allocateBoneSet();
+            if (instance.boneSet[frameIndex]) {
+                VkDescriptorBufferInfo bufInfo{};
+                bufInfo.buffer = instance.boneBuffer[frameIndex];
+                bufInfo.offset = 0;
+                bufInfo.range = bci.size;
+                VkWriteDescriptorSet write{VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET};
+                write.dstSet = instance.boneSet[frameIndex];
+                write.dstBinding = 0;
+                write.descriptorCount = 1;
+                write.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
+                write.pBufferInfo = &bufInfo;
+                vkUpdateDescriptorSets(vkCtx_->getDevice(), 1, &write, 0, nullptr);
+            }
+        }
+    }
+}
+
 void M2Renderer::render(VkCommandBuffer cmd, VkDescriptorSet perFrameSet, const Camera& camera) {
     if (instances.empty() || !opaquePipeline_) {
         return;
@@ -2254,7 +2303,7 @@ void M2Renderer::render(VkCommandBuffer cmd, VkDescriptorSet perFrameSet, const
     }
 
     // Sort by modelId to minimize vertex/index buffer rebinds
-    std::stable_sort(sortedVisible_.begin(), sortedVisible_.end(),
+    std::sort(sortedVisible_.begin(), sortedVisible_.end(),
               [](const VisibleEntry& a, const VisibleEntry& b) { return a.modelId < b.modelId; });
 
     uint32_t currentModelId = UINT32_MAX;
@@ -2330,44 +2379,22 @@ void M2Renderer::render(VkCommandBuffer cmd, VkDescriptorSet perFrameSet, const
             }
         }
 
-        // Upload bone matrices to SSBO if model has skeletal animation
-        bool useBones = model.hasAnimation && !model.disableAnimation && !instance.boneMatrices.empty();
+        // Upload bone matrices to SSBO if model has skeletal animation.
+        // Bone buffers are pre-allocated by prepareRender() on the main thread.
+        // If not yet allocated (race/timing), skip this instance entirely to avoid
+        // a bind-pose flash — it will render correctly next frame.
+        bool needsBones = model.hasAnimation && !model.disableAnimation && !instance.boneMatrices.empty();
+        if (needsBones && (!instance.boneBuffer[frameIndex] || !instance.boneSet[frameIndex])) {
+            continue;
+        }
+        bool useBones = needsBones;
         if (useBones) {
-            // Lazy-allocate bone SSBO on first use
-            if (!instance.boneBuffer[frameIndex]) {
-                VkBufferCreateInfo bci{VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO};
-                bci.size = 128 * sizeof(glm::mat4); // max 128 bones
-                bci.usage = VK_BUFFER_USAGE_STORAGE_BUFFER_BIT;
-                VmaAllocationCreateInfo aci{};
-                aci.usage = VMA_MEMORY_USAGE_CPU_TO_GPU;
-                aci.flags = VMA_ALLOCATION_CREATE_MAPPED_BIT;
-                VmaAllocationInfo allocInfo{};
-                vmaCreateBuffer(vkCtx_->getAllocator(), &bci, &aci,
-                                &instance.boneBuffer[frameIndex], &instance.boneAlloc[frameIndex], &allocInfo);
-                instance.boneMapped[frameIndex] = allocInfo.pMappedData;
-
-                // Allocate descriptor set for bone SSBO
-                instance.boneSet[frameIndex] = allocateBoneSet();
-                if (instance.boneSet[frameIndex]) {
-                    VkDescriptorBufferInfo bufInfo{};
-                    bufInfo.buffer = instance.boneBuffer[frameIndex];
-                    bufInfo.offset = 0;
-                    bufInfo.range = bci.size;
-                    VkWriteDescriptorSet write{VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET};
-                    write.dstSet = instance.boneSet[frameIndex];
-                    write.dstBinding = 0;
-                    write.descriptorCount = 1;
-                    write.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
-                    write.pBufferInfo = &bufInfo;
-                    vkUpdateDescriptorSets(vkCtx_->getDevice(), 1, &write, 0, nullptr);
-                }
-            }
-
-            // Upload bone matrices
-            if (instance.boneMapped[frameIndex]) {
+            // Upload bone matrices only when recomputed (skip frame-skipped instances)
+            if (instance.bonesDirty && instance.boneMapped[frameIndex]) {
                 int numBones = std::min(static_cast<int>(instance.boneMatrices.size()), 128);
                 memcpy(instance.boneMapped[frameIndex], instance.boneMatrices.data(),
                        numBones * sizeof(glm::mat4));
+                instance.bonesDirty = false;
             }
 
             // Bind bone descriptor set (set 2)
@@ -2384,12 +2411,8 @@ void M2Renderer::render(VkCommandBuffer cmd, VkDescriptorSet perFrameSet, const
         else if (entry.distSq > 40.0f * 40.0f) desiredLOD = 1;
 
         uint16_t targetLOD = desiredLOD;
-        if (desiredLOD > 0) {
-            bool hasDesiredLOD = false;
-            for (const auto& b : model.batches) {
-                if (b.submeshLevel == desiredLOD) { hasDesiredLOD = true; break; }
-            }
-            if (!hasDesiredLOD) targetLOD = 0;
+        if (desiredLOD > 0 && !(model.availableLODs & (1u << desiredLOD))) {
+            targetLOD = 0;
         }
 
         const bool foliageLikeModel = model.isFoliageLike;

src/rendering/performance_hud.cpp

@@ -1,5 +1,6 @@
 #include "rendering/performance_hud.hpp"
 #include "rendering/renderer.hpp"
+#include "rendering/vk_context.hpp"
 #include "rendering/terrain_renderer.hpp"
 #include "rendering/terrain_manager.hpp"
 #include "rendering/water_renderer.hpp"
@@ -187,6 +188,19 @@ void PerformanceHUD::render(const Renderer* renderer, const Camera* camera) {
                            0, nullptr, 0.0f, 33.33f, ImVec2(200, 40));
         }
 
+        // FSR info
+        if (renderer->isFSREnabled()) {
+            ImGui::TextColored(ImVec4(0.4f, 1.0f, 0.4f, 1.0f), "FSR 1.0: ON");
+            auto* ctx = renderer->getVkContext();
+            if (ctx) {
+                auto ext = ctx->getSwapchainExtent();
+                float sf = renderer->getFSRScaleFactor();
+                uint32_t iw = static_cast<uint32_t>(ext.width * sf) & ~1u;
+                uint32_t ih = static_cast<uint32_t>(ext.height * sf) & ~1u;
+                ImGui::Text("  %ux%u -> %ux%u (%.0f%%)", iw, ih, ext.width, ext.height, sf * 100.0f);
+            }
+        }
+
         ImGui::Spacing();
     }
 

src/rendering/terrain_manager.cpp

@@ -911,6 +911,8 @@ bool TerrainManager::advanceFinalization(FinalizingTile& ft) {
             wmoRenderer->setDeferNormalMaps(false);
             wmoRenderer->setPredecodedBLPCache(nullptr);
             if (ft.wmoModelIndex < pending->wmoModels.size()) return false;
+            // All WMO models loaded — backfill normal/height maps that were skipped during streaming
+            wmoRenderer->backfillNormalMaps();
         }
         ft.phase = FinalizationPhase::WMO_INSTANCES;
         return false;

src/rendering/vk_context.cpp

@@ -252,14 +252,22 @@ bool VkContext::createAllocator() {
 bool VkContext::createSwapchain(int width, int height) {
     vkb::SwapchainBuilder swapchainBuilder{physicalDevice, device, surface};
 
-    auto swapRet = swapchainBuilder
+    auto& builder = swapchainBuilder
         .set_desired_format({VK_FORMAT_B8G8R8A8_UNORM, VK_COLOR_SPACE_SRGB_NONLINEAR_KHR})
-        .set_desired_present_mode(VK_PRESENT_MODE_FIFO_KHR) // VSync
         .set_desired_extent(static_cast<uint32_t>(width), static_cast<uint32_t>(height))
         .set_image_usage_flags(VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT)
         .set_desired_min_image_count(2)
-        .set_old_swapchain(swapchain) // For recreation
-        .build();
+        .set_old_swapchain(swapchain);
+
+    if (vsync_) {
+        builder.set_desired_present_mode(VK_PRESENT_MODE_FIFO_KHR);
+    } else {
+        builder.set_desired_present_mode(VK_PRESENT_MODE_IMMEDIATE_KHR);
+        builder.add_fallback_present_mode(VK_PRESENT_MODE_MAILBOX_KHR);
+        builder.add_fallback_present_mode(VK_PRESENT_MODE_FIFO_RELAXED_KHR);
+    }
+
+    auto swapRet = builder.build();
 
     if (!swapRet) {
         LOG_ERROR("Failed to create Vulkan swapchain: ", swapRet.error().message());
@@ -1026,14 +1034,22 @@ bool VkContext::recreateSwapchain(int width, int height) {
     VkSwapchainKHR oldSwapchain = swapchain;
 
     vkb::SwapchainBuilder swapchainBuilder{physicalDevice, device, surface};
-    auto swapRet = swapchainBuilder
+    auto& builder = swapchainBuilder
         .set_desired_format({VK_FORMAT_B8G8R8A8_UNORM, VK_COLOR_SPACE_SRGB_NONLINEAR_KHR})
-        .set_desired_present_mode(VK_PRESENT_MODE_FIFO_KHR)
         .set_desired_extent(static_cast<uint32_t>(width), static_cast<uint32_t>(height))
         .set_image_usage_flags(VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT)
         .set_desired_min_image_count(2)
-        .set_old_swapchain(oldSwapchain)
-        .build();
+        .set_old_swapchain(oldSwapchain);
+
+    if (vsync_) {
+        builder.set_desired_present_mode(VK_PRESENT_MODE_FIFO_KHR);
+    } else {
+        builder.set_desired_present_mode(VK_PRESENT_MODE_IMMEDIATE_KHR);
+        builder.add_fallback_present_mode(VK_PRESENT_MODE_MAILBOX_KHR);
+        builder.add_fallback_present_mode(VK_PRESENT_MODE_FIFO_RELAXED_KHR);
+    }
+
+    auto swapRet = builder.build();
 
     if (oldSwapchain) {
         vkDestroySwapchainKHR(device, oldSwapchain, nullptr);

src/rendering/wmo_renderer.cpp

@@ -787,8 +787,8 @@ bool WMORenderer::loadModel(const pipeline::WMOModel& model, uint32_t id) {
             }
 
             // Build doodad's local transform (WoW coordinates)
-            // WMO doodads use quaternion rotation (X/Y swapped for correct orientation)
-            glm::quat fixedRotation(doodad.rotation.w, doodad.rotation.y, doodad.rotation.x, doodad.rotation.z);
+            // WMO doodads use quaternion rotation
+            glm::quat fixedRotation(doodad.rotation.w, doodad.rotation.x, doodad.rotation.y, doodad.rotation.z);
 
             glm::mat4 localTransform(1.0f);
             localTransform = glm::translate(localTransform, doodad.position);
@@ -1318,15 +1318,10 @@ void WMORenderer::gatherCandidates(const glm::vec3& queryMin, const glm::vec3& q
     }
 }
 
-void WMORenderer::render(VkCommandBuffer cmd, VkDescriptorSet perFrameSet, const Camera& camera) {
+void WMORenderer::prepareRender() {
     ++currentFrameId;
 
-    if (!opaquePipeline_ || instances.empty()) {
-        lastDrawCalls = 0;
-        return;
-    }
-
-    // Update material UBOs if settings changed
+    // Update material UBOs if settings changed (mapped memory writes — main thread only)
     if (materialSettingsDirty_) {
         materialSettingsDirty_ = false;
         static const int pomSampleTable[] = { 16, 32, 64 };
@@ -1335,7 +1330,6 @@ void WMORenderer::render(VkCommandBuffer cmd, VkDescriptorSet perFrameSet, const
             for (auto& group : model.groups) {
                 for (auto& mb : group.mergedBatches) {
                     if (!mb.materialUBO) continue;
-                    // Read existing UBO data, update normal/POM fields
                     VmaAllocationInfo allocInfo{};
                     vmaGetAllocationInfo(vkCtx_->getAllocator(), mb.materialUBOAlloc, &allocInfo);
                     if (allocInfo.pMappedData) {
@@ -1351,6 +1345,13 @@ void WMORenderer::render(VkCommandBuffer cmd, VkDescriptorSet perFrameSet, const
             }
         }
     }
+}
+
+void WMORenderer::render(VkCommandBuffer cmd, VkDescriptorSet perFrameSet, const Camera& camera) {
+    if (!opaquePipeline_ || instances.empty()) {
+        lastDrawCalls = 0;
+        return;
+    }
 
     lastDrawCalls = 0;
 
@@ -1362,43 +1363,45 @@ void WMORenderer::render(VkCommandBuffer cmd, VkDescriptorSet perFrameSet, const
     lastPortalCulledGroups = 0;
     lastDistanceCulledGroups = 0;
 
-    // ── Phase 1: Parallel visibility culling ──────────────────────────
-    std::vector<size_t> visibleInstances;
-    visibleInstances.reserve(instances.size());
+    // ── Phase 1: Visibility culling ──────────────────────────
+    visibleInstances_.clear();
     for (size_t i = 0; i < instances.size(); ++i) {
-        const auto& instance = instances[i];
-        if (loadedModels.find(instance.modelId) == loadedModels.end())
-            continue;
-        visibleInstances.push_back(i);
+        if (loadedModels.count(instances[i].modelId))
+            visibleInstances_.push_back(i);
     }
 
     glm::vec3 camPos = camera.getPosition();
     bool doPortalCull = portalCulling;
-    bool doFrustumCull = false; // Temporarily disabled: can over-cull world WMOs
     bool doDistanceCull = distanceCulling;
 
-    auto cullInstance = [&](size_t instIdx) -> InstanceDrawList {
-        if (instIdx >= instances.size()) return InstanceDrawList{};
+    auto cullInstance = [&](size_t instIdx, InstanceDrawList& result) {
+        if (instIdx >= instances.size()) return;
         const auto& instance = instances[instIdx];
         auto mdlIt = loadedModels.find(instance.modelId);
-        if (mdlIt == loadedModels.end()) return InstanceDrawList{};
+        if (mdlIt == loadedModels.end()) return;
         const ModelData& model = mdlIt->second;
 
-        InstanceDrawList result;
         result.instanceIndex = instIdx;
+        result.visibleGroups.clear();
+        result.portalCulled = 0;
+        result.distanceCulled = 0;
 
-        // Portal-based visibility
-        std::unordered_set<uint32_t> portalVisibleGroups;
+        // Portal-based visibility — use a flat sorted vector instead of unordered_set
+        std::vector<uint32_t> portalVisibleGroups;
         bool usePortalCulling = doPortalCull && !model.portals.empty() && !model.portalRefs.empty();
         if (usePortalCulling) {
+            std::unordered_set<uint32_t> pvgSet;
             glm::vec4 localCamPos = instance.invModelMatrix * glm::vec4(camPos, 1.0f);
             getVisibleGroupsViaPortals(model, glm::vec3(localCamPos), frustum,
-                                       instance.modelMatrix, portalVisibleGroups);
+                                       instance.modelMatrix, pvgSet);
+            portalVisibleGroups.assign(pvgSet.begin(), pvgSet.end());
+            std::sort(portalVisibleGroups.begin(), portalVisibleGroups.end());
         }
 
         for (size_t gi = 0; gi < model.groups.size(); ++gi) {
             if (usePortalCulling &&
-                portalVisibleGroups.find(static_cast<uint32_t>(gi)) == portalVisibleGroups.end()) {
+                !std::binary_search(portalVisibleGroups.begin(), portalVisibleGroups.end(),
+                                    static_cast<uint32_t>(gi))) {
                 result.portalCulled++;
                 continue;
             }
@@ -1414,62 +1417,18 @@ void WMORenderer::render(VkCommandBuffer cmd, VkDescriptorSet perFrameSet, const
                         continue;
                     }
                 }
-
-                if (doFrustumCull && !frustum.intersectsAABB(gMin, gMax))
-                    continue;
             }
 
             result.visibleGroups.push_back(static_cast<uint32_t>(gi));
         }
-        return result;
     };
 
-    // Dispatch culling — parallel when enough instances, sequential otherwise.
-    std::vector<InstanceDrawList> drawLists;
-    drawLists.reserve(visibleInstances.size());
+    // Resize drawLists to match (reuses previous capacity)
+    drawLists_.resize(visibleInstances_.size());
 
-    static const size_t minParallelCullInstances = std::max<size_t>(
-        4, envSizeOrDefault("WOWEE_WMO_CULL_MT_MIN", 128));
-    if (visibleInstances.size() >= minParallelCullInstances && numCullThreads_ > 1) {
-        static const size_t minCullWorkPerThread = std::max<size_t>(
-            16, envSizeOrDefault("WOWEE_WMO_CULL_WORK_PER_THREAD", 64));
-        const size_t maxUsefulThreads = std::max<size_t>(
-            1, (visibleInstances.size() + minCullWorkPerThread - 1) / minCullWorkPerThread);
-        const size_t numThreads = std::min(static_cast<size_t>(numCullThreads_), maxUsefulThreads);
-        if (numThreads <= 1) {
-            for (size_t idx : visibleInstances) {
-                drawLists.push_back(cullInstance(idx));
-            }
-        } else {
-            const size_t chunkSize = visibleInstances.size() / numThreads;
-            const size_t remainder = visibleInstances.size() % numThreads;
-
-            drawLists.resize(visibleInstances.size());
-
-            cullFutures_.clear();
-            if (cullFutures_.capacity() < numThreads) {
-                cullFutures_.reserve(numThreads);
-            }
-
-            size_t start = 0;
-            for (size_t t = 0; t < numThreads; ++t) {
-                const size_t end = start + chunkSize + (t < remainder ? 1 : 0);
-                cullFutures_.push_back(std::async(std::launch::async,
-                    [&, start, end]() {
-                        for (size_t j = start; j < end; ++j) {
-                            drawLists[j] = cullInstance(visibleInstances[j]);
-                        }
-                    }));
-                start = end;
-            }
-
-            for (auto& f : cullFutures_) {
-                f.get();
-            }
-        }
-    } else {
-        for (size_t idx : visibleInstances)
-            drawLists.push_back(cullInstance(idx));
+    // Sequential culling (parallel dispatch overhead > savings for typical instance counts)
+    for (size_t j = 0; j < visibleInstances_.size(); ++j) {
+        cullInstance(visibleInstances_[j], drawLists_[j]);
     }
 
     // ── Phase 2: Vulkan draw ────────────────────────────────
@@ -1484,7 +1443,7 @@ void WMORenderer::render(VkCommandBuffer cmd, VkDescriptorSet perFrameSet, const
     // Track which pipeline is currently bound: 0=opaque, 1=transparent, 2=glass
     int currentPipelineKind = 0;
 
-    for (const auto& dl : drawLists) {
+    for (const auto& dl : drawLists_) {
         if (dl.instanceIndex >= instances.size()) continue;
         const auto& instance = instances[dl.instanceIndex];
         auto modelIt = loadedModels.find(instance.modelId);
@@ -2412,6 +2371,69 @@ VkTexture* WMORenderer::loadTexture(const std::string& path) {
     return rawPtr;
 }
 
+void WMORenderer::backfillNormalMaps() {
+    if (!normalMappingEnabled_ && !pomEnabled_) return;
+
+    if (!assetManager) return;
+
+    int generated = 0;
+    for (auto& [key, entry] : textureCache) {
+        if (entry.normalHeightMap) continue;  // already has one
+        if (!entry.texture) continue;
+
+        // Re-load the BLP from MPQ to get pixel data for normal map generation
+        pipeline::BLPImage blp = assetManager->loadTexture(key);
+        if (!blp.isValid() || blp.width == 0 || blp.height == 0) continue;
+
+        float variance = 0.0f;
+        auto nhMap = generateNormalHeightMap(blp.data.data(), blp.width, blp.height, variance);
+        if (nhMap) {
+            entry.normalHeightMap = std::move(nhMap);
+            entry.heightMapVariance = variance;
+            generated++;
+        }
+    }
+
+    if (generated > 0) {
+        VkDevice device = vkCtx_->getDevice();
+        int rebound = 0;
+        // Update merged batches: assign normal map pointer and rebind descriptor set
+        for (auto& [modelId, model] : loadedModels) {
+            for (auto& group : model.groups) {
+                for (auto& mb : group.mergedBatches) {
+                    if (mb.normalHeightMap) continue;  // already set
+                    if (!mb.texture) continue;
+                    // Find this texture in the cache
+                    for (const auto& [cacheKey, cacheEntry] : textureCache) {
+                        if (cacheEntry.texture.get() == mb.texture) {
+                            if (cacheEntry.normalHeightMap) {
+                                mb.normalHeightMap = cacheEntry.normalHeightMap.get();
+                                mb.heightMapVariance = cacheEntry.heightMapVariance;
+                                // Rebind descriptor set binding 2 to the real normal/height map
+                                if (mb.materialSet) {
+                                    VkDescriptorImageInfo nhImgInfo = mb.normalHeightMap->descriptorInfo();
+                                    VkWriteDescriptorSet write{};
+                                    write.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
+                                    write.dstSet = mb.materialSet;
+                                    write.dstBinding = 2;
+                                    write.descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
+                                    write.descriptorCount = 1;
+                                    write.pImageInfo = &nhImgInfo;
+                                    vkUpdateDescriptorSets(device, 1, &write, 0, nullptr);
+                                    rebound++;
+                                }
+                            }
+                            break;
+                        }
+                    }
+                }
+            }
+        }
+        materialSettingsDirty_ = true;
+        LOG_INFO("Backfilled ", generated, " normal/height maps (", rebound, " descriptor sets rebound) for deferred WMO textures");
+    }
+}
+
 // Ray-AABB intersection (slab method)
 // Returns true if the ray intersects the axis-aligned bounding box
 static bool rayIntersectsAABB(const glm::vec3& origin, const glm::vec3& dir,
@@ -3145,18 +3167,13 @@ bool WMORenderer::checkWallCollision(const glm::vec3& from, const glm::vec3& to,
                 if (triHeight < 1.0f && tb.maxZ <= localFeetZ + 1.2f) continue;
 
                 // Use MOPY flags to filter wall collision.
-                // Collidable triangles (flag 0x01) block the player — including
-                // invisible collision walls (0x01 without 0x20) used in tunnels.
-                // Skip detail/decorative geometry (0x04) and render-only surfaces.
+                // Collide with triangles that have the collision flag (0x08) or no flags at all.
+                // Skip detail/decorative (0x04) and render-only (0x20 without 0x08) surfaces.
                 uint32_t triIdx = triStart / 3;
                 if (!group.triMopyFlags.empty() && triIdx < group.triMopyFlags.size()) {
                     uint8_t mopy = group.triMopyFlags[triIdx];
                     if (mopy != 0) {
-                        bool collidable = (mopy & 0x01) != 0;
-                        bool detail = (mopy & 0x04) != 0;
-                        if (!collidable || detail) {
-                            continue;
-                        }
+                        if ((mopy & 0x04) || !(mopy & 0x08)) continue;
                     }
                 }
 
@@ -3217,8 +3234,8 @@ bool WMORenderer::checkWallCollision(const glm::vec3& from, const glm::vec3& to,
                     if (absNz >= 0.35f) continue;
 
                     const float SKIN = 0.005f;        // small separation so we don't re-collide immediately
-                    // Stronger push when inside WMO for more responsive indoor collision
-                    const float MAX_PUSH = insideWMO ? 0.35f : 0.15f;
+                    // Push must cover full penetration to prevent gradual clip-through
+                    const float MAX_PUSH = PLAYER_RADIUS;
                     float penetration = (PLAYER_RADIUS - horizDist);
                     float pushDist = glm::clamp(penetration + SKIN, 0.0f, MAX_PUSH);
                     glm::vec2 pushDir2;

src/ui/game_screen.cpp

@@ -317,6 +317,20 @@ void GameScreen::render(game::GameHandler& gameHandler) {
         }
     }
 
+    // Apply saved FSR setting once when renderer is available
+    if (!fsrSettingsApplied_ && pendingFSR) {
+        auto* renderer = core::Application::getInstance().getRenderer();
+        if (renderer) {
+            static const float fsrScales[] = { 0.77f, 0.67f, 0.59f, 0.50f };
+            renderer->setFSRQuality(fsrScales[pendingFSRQuality]);
+            renderer->setFSRSharpness(pendingFSRSharpness);
+            renderer->setFSREnabled(true);
+            fsrSettingsApplied_ = true;
+        }
+    } else {
+        fsrSettingsApplied_ = true;
+    }
+
     // Apply auto-loot setting to GameHandler every frame (cheap bool sync)
     gameHandler.setAutoLoot(pendingAutoLoot);
 
@@ -2687,6 +2701,12 @@ void GameScreen::sendChatMessage(game::GameHandler& gameHandler) {
                 chatInputBuffer[0] = '\0';
                 return;
             }
+            // /unstuckhearth command — teleport to hearthstone bind point
+            if (cmdLower == "unstuckhearth") {
+                gameHandler.unstuckHearth();
+                chatInputBuffer[0] = '\0';
+                return;
+            }
 
             // /transport board — board test transport
             if (cmdLower == "transport board") {
@@ -6270,6 +6290,25 @@ void GameScreen::renderSettingsWindow() {
                         saveSettings();
                     }
                 }
+                // FSR 1.0 Upscaling
+                {
+                    if (ImGui::Checkbox("FSR Upscaling (Experimental)", &pendingFSR)) {
+                        if (renderer) renderer->setFSREnabled(pendingFSR);
+                        saveSettings();
+                    }
+                    if (pendingFSR) {
+                        const char* fsrQualityLabels[] = { "Ultra Quality (77%)", "Quality (67%)", "Balanced (59%)", "Performance (50%)" };
+                        static const float fsrScaleFactors[] = { 0.77f, 0.67f, 0.59f, 0.50f };
+                        if (ImGui::Combo("FSR Quality", &pendingFSRQuality, fsrQualityLabels, 4)) {
+                            if (renderer) renderer->setFSRQuality(fsrScaleFactors[pendingFSRQuality]);
+                            saveSettings();
+                        }
+                        if (ImGui::SliderFloat("FSR Sharpness", &pendingFSRSharpness, 0.0f, 2.0f, "%.1f")) {
+                            if (renderer) renderer->setFSRSharpness(pendingFSRSharpness);
+                            saveSettings();
+                        }
+                    }
+                }
                 if (ImGui::SliderInt("Ground Clutter Density", &pendingGroundClutterDensity, 0, 150, "%d%%")) {
                     if (renderer) {
                         if (auto* tm = renderer->getTerrainManager()) {
@@ -7384,6 +7423,9 @@ void GameScreen::saveSettings() {
     out << "normal_map_strength=" << pendingNormalMapStrength << "\n";
     out << "pom=" << (pendingPOM ? 1 : 0) << "\n";
     out << "pom_quality=" << pendingPOMQuality << "\n";
+    out << "fsr=" << (pendingFSR ? 1 : 0) << "\n";
+    out << "fsr_quality=" << pendingFSRQuality << "\n";
+    out << "fsr_sharpness=" << pendingFSRSharpness << "\n";
 
     // Controls
     out << "mouse_sensitivity=" << pendingMouseSensitivity << "\n";
@@ -7470,6 +7512,9 @@ void GameScreen::loadSettings() {
             else if (key == "normal_map_strength") pendingNormalMapStrength = std::clamp(std::stof(val), 0.0f, 2.0f);
             else if (key == "pom") pendingPOM = (std::stoi(val) != 0);
             else if (key == "pom_quality") pendingPOMQuality = std::clamp(std::stoi(val), 0, 2);
+            else if (key == "fsr") pendingFSR = (std::stoi(val) != 0);
+            else if (key == "fsr_quality") pendingFSRQuality = std::clamp(std::stoi(val), 0, 3);
+            else if (key == "fsr_sharpness") pendingFSRSharpness = std::clamp(std::stof(val), 0.0f, 2.0f);
             // Controls
             else if (key == "mouse_sensitivity") pendingMouseSensitivity = std::clamp(std::stof(val), 0.05f, 1.0f);
             else if (key == "invert_mouse") pendingInvertMouse = (std::stoi(val) != 0);