diff --git a/mod/docs/optimize-layer-design.md b/mod/docs/optimize-layer-design.md
index f240bf10..56609205 100644
--- a/mod/docs/optimize-layer-design.md
+++ b/mod/docs/optimize-layer-design.md
@@ -3,6 +3,12 @@
 Статус: Этапы 1-3 готовы (2026-10-12). Этап 4 (Vulkan) — стартуем с универсально-полезных пунктов. Цель MC: 26.2.
 
 ## История изменений
+- **2026-10-12**: Этап 4 старт — `RenderBackendOps.current()` теперь определяет активный бэкенд через
+  `RenderSystem.getDevice().getClass().getName()` (class-name, не `instanceof`, чтобы не форсировать линковку
+  `GlDevice`/`VulkanDevice`); добавлен `misc/optimize/vulkan/VulkanRenderBackendOps` — no-op-заглушка со
+  всеми `*Supported() == false`. Под GL всё работает как раньше; под Vulkan mixins из `mixins/optimize/gl/*` не
+  загружаются, и теперь хотя бы интерфейс честно возвращает правильный бэкенд. Живые Vulkan-пути
+  (`VkPipelineCache`, state cache) — следующий шаг.
 - **2026-10-12**: убран гейт «меряем на Windows до включения». Бенчмарки больше не блокируют работу: универсально-полезные
   GL/Vulkan-оптимизации делаем сразу за default-off тумблерами (пользователь включает сам), отчёт цифр остаётся как
   факультативная задача владельца и не входит в план. Этап 3 закрыт по всем пунктам, кроме п.9 (`no_item_sort` /
diff --git a/mod/src/main/java/dev/loki/lovisual/features/module/modules/misc/optimize/gl/RenderBackendOps.java b/mod/src/main/java/dev/loki/lovisual/features/module/modules/misc/optimize/gl/RenderBackendOps.java
index 00605c8f..3f2feedf 100644
--- a/mod/src/main/java/dev/loki/lovisual/features/module/modules/misc/optimize/gl/RenderBackendOps.java
+++ b/mod/src/main/java/dev/loki/lovisual/features/module/modules/misc/optimize/gl/RenderBackendOps.java
@@ -1,5 +1,8 @@
 package dev.loki.lovisual.features.module.modules.misc.optimize.gl;
 
+import com.mojang.blaze3d.systems.RenderSystem;
+import dev.loki.lovisual.features.module.modules.misc.optimize.vulkan.VulkanRenderBackendOps;
+
 /**
  * Backend-agnostic surface for the optimization layer's GL/Vulkan hooks.
  *
@@ -29,8 +32,30 @@ package dev.loki.lovisual.features.module.modules.misc.optimize.gl;
  */
 public interface RenderBackendOps {
 
-    /** Resolve the current backend — GL backend if GL is active, otherwise {@code null}. */
+    /**
+     * Resolve the current backend by inspecting the active {@code GpuDeviceBackend}.
+     * Class-name check (not {@code instanceof}) so this method itself does not
+     * force the loader to link {@code GlDevice} or {@code VulkanDevice}; the
+     * Blaze3D jar always ships both, but keeping the check by string means the
+     * dispatch stays correct even when only one backend's class has been
+     * resolved. Falls back to the GL impl when the device is null (early boot)
+     * so mixins' {@code supported()} short-circuits behave the same as before.
+     */
     static RenderBackendOps current() {
+        Object device;
+        try {
+            device = RenderSystem.getDevice();
+        } catch (Throwable ignored) {
+            // Tests and pre-init paths call current() before RenderSystem has a
+            // device; fall through to GL so existing unit tests keep passing.
+            device = null;
+        }
+        if (device != null) {
+            String name = device.getClass().getName();
+            if (name.startsWith("com.mojang.blaze3d.vulkan.")) {
+                return VulkanRenderBackendOps.current();
+            }
+        }
         return GlRenderBackendOps.current();
     }
 
diff --git a/mod/src/main/java/dev/loki/lovisual/features/module/modules/misc/optimize/vulkan/VulkanRenderBackendOps.java b/mod/src/main/java/dev/loki/lovisual/features/module/modules/misc/optimize/vulkan/VulkanRenderBackendOps.java
new file mode 100644
index 00000000..fccf117a
--- /dev/null
+++ b/mod/src/main/java/dev/loki/lovisual/features/module/modules/misc/optimize/vulkan/VulkanRenderBackendOps.java
@@ -0,0 +1,97 @@
+package dev.loki.lovisual.features.module.modules.misc.optimize.vulkan;
+
+import dev.loki.lovisual.features.module.modules.misc.optimize.gl.RenderBackendOps;
+
+/**
+ * Vulkan-side implementation of {@link RenderBackendOps}. All GPU-object ops
+ * (lazy clear, TBO cache, shared depth FBO, RenderTypes cache) currently live
+ * in the GL mixin set — under Vulkan those mixins never load, so this class is
+ * a no-op placeholder that lets {@code RenderBackendOps.current()} dispatch by
+ * actual backend instead of unconditionally returning the GL impl.
+ *
+ * <p>Every {@code *Supported} reports {@code false}. Callers already gate on
+ * supported before attempting, so the effect is the same as "Optimize GPU
+ * features off" for backend-dependent items. Stage 4 adds Vulkan-specific
+ * paths (persistent {@code VkPipelineCache}, {@code VulkanRenderPass} state
+ * cache) by flipping the corresponding {@code *Supported} here to true and
+ * implementing the op body.
+ *
+ * <p>Doc reference: {@code mod/docs/optimize-layer-design.md}, "Этап 4" and
+ * "Что уже решено в архитектуре" → "Рендер-бэкенды".
+ */
+public final class VulkanRenderBackendOps implements RenderBackendOps {
+
+    private static final VulkanRenderBackendOps INSTANCE = new VulkanRenderBackendOps();
+
+    private VulkanRenderBackendOps() {
+    }
+
+    /** Singleton — no state, no per-device instance needed. */
+    public static VulkanRenderBackendOps current() {
+        return INSTANCE;
+    }
+
+    // --- Lazy clear ---
+
+    @Override
+    public boolean lazyClearSupported() {
+        return false;
+    }
+
+    @Override
+    public boolean skipBindFramebufferZero(int framebuffer) {
+        return false;
+    }
+
+    // --- TBO cache ---
+
+    @Override
+    public boolean tboCacheSupported() {
+        return false;
+    }
+
+    @Override
+    public void trackBoundTboTexture(int texture) {
+    }
+
+    @Override
+    public boolean tboAlreadyAttached(int format, int buffer) {
+        return false;
+    }
+
+    @Override
+    public void markTboAttached(int format, int buffer) {
+    }
+
+    @Override
+    public void forgetTboCache() {
+    }
+
+    // --- FBO share ---
+
+    @Override
+    public boolean fboShareSupported() {
+        return false;
+    }
+
+    @Override
+    public boolean clearDepthInPlace(int depthId, double value) {
+        return false;
+    }
+
+    // --- RenderTypes cache ---
+
+    @Override
+    public boolean renderTypesCacheSupported() {
+        return false;
+    }
+
+    @Override
+    public Object renderTypeFor(Identifier id) {
+        return null;
+    }
+
+    @Override
+    public void cacheRenderType(Identifier id, Object type) {
+    }
+}