Use software decoder by default

modules/ANSODEngine/engine.h

@@ -402,6 +402,9 @@ private:
     cudaStream_t m_memoryStream;  // ADD THIS - separate stream for memory operations
     std::vector<cv::cuda::GpuMat> m_preprocessedInputs;  // Keep inputs alive
 
+    // Note: blobFromGpuMats and resizeKeepAspectRatioPadRightBottom are static,
+    // so cached buffers use thread_local inside the functions themselves.
+
 
     // Thermal management (ADD THESE)
     //int m_consecutiveInferences;
@@ -431,7 +434,7 @@ private:
 
     Logger m_logger;
     bool m_verbose{ true };   // false for non-probe pool slots
-    bool m_disableGraphs{ false }; // true for pool slots — concurrent graph captures corrupt CUDA context
+    bool m_disableGraphs{ true }; // DISABLED by default — concurrent graph launches + uploads cause GPU deadlock on WDDM
 
     // -- Multi-GPU pool data ---------------------------------------------------