Disable NV12 path for ANSCV by default. Currenly use cv::Mat** directly

include/ANSGpuFrameRegistry.h

@@ -132,6 +132,13 @@ struct GpuFrameData {
     // freed while any consumer is still reading it.
     GpuNV12Slot* poolSlot = nullptr;
 
+    // --- Async D2D copy stream ---
+    // The CUDA stream used for the async D2D copy from NVDEC surface to pool buffer.
+    // Inference MUST call cudaStreamSynchronize on this before reading yPlane/uvPlane
+    // to ensure the copy has completed.  Stored as void* to avoid cuda_runtime.h here.
+    // nullptr means D2D was synchronous (legacy path) or no D2D copy was done.
+    void* d2dCopyStream = nullptr;
+
     // Default constructor
     GpuFrameData() = default;
 
@@ -151,6 +158,7 @@ struct GpuFrameData {
         , refcount(o.refcount.load()), createdAt(o.createdAt)
         , ownerClient(o.ownerClient), onReleaseFn(o.onReleaseFn)
         , poolSlot(o.poolSlot)
+        , d2dCopyStream(o.d2dCopyStream)
     {
         // Null out source to prevent double-free of owned pointers
         o.cpuYPlane = nullptr;
@@ -165,6 +173,7 @@ struct GpuFrameData {
         o.ownerClient = nullptr;
         o.onReleaseFn = nullptr;
         o.poolSlot = nullptr;
+        o.d2dCopyStream = nullptr;
     }
 
     // No copy
@@ -360,6 +369,12 @@ public:
         return result;
     }
 
+    // Push an AVFrame* (as void*) for deferred freeing.
+    // Caller MUST hold the lock via acquire_lock().
+    void pushPendingFree_locked(void* ptr) {
+        if (ptr) m_pendingFree.push_back(ptr);
+    }
+
     // --- Drain pending GPU device pointers for caller to cudaFree ---
     // Each entry includes the device index for cudaSetDevice before cudaFree.
     // If minAgeMs > 0, only drain entries older than minAgeMs milliseconds.