Use software decoder by default

tests/ANSLPR-UnitTest/ANSLPR-UnitTest.cpp

@@ -814,8 +814,8 @@ static void ALPRWorkerThread(int taskId,
                 g_log.add(prefix + " Empty frame (count=" + std::to_string(emptyFrames) + ")");
             }
             if (emptyFrames > 300) {
-                g_log.add(prefix + " Too many empty frames, attempting reconnect...");
-                ReconnectRTSP(&rtspClient);
+                g_log.add(prefix + " Too many empty frames (reconnect disabled for long test)");
+                // ReconnectRTSP(&rtspClient);  // Disabled for VRAM stability testing
                 emptyFrames = 0;
             }
             streamLock.unlock();
@@ -1222,9 +1222,9 @@ int ANSLPR_MultiGPU_StressTest() {
                     g_log.add(buf);
                     printf("%s\n", buf);
                 } else if (currentGpu != streamPreferredGpu[s]) {
-                    // Decoder is active on wrong GPU — reconnect to move it
+                    // Decoder is active on wrong GPU — reconnect disabled for VRAM stability testing
                     SetRTSPHWDecoding(&rtspClients[s], 7, streamPreferredGpu[s]);
-                    ReconnectRTSP(&rtspClients[s]);
+                    // ReconnectRTSP(&rtspClients[s]);  // Disabled for long test
                     char buf[256];
                     snprintf(buf, sizeof(buf),
                         "[Stream%d] NVDEC GPU realigned: GPU[%d] -> GPU[%d] (reconnected for zero-copy)",
@@ -1279,7 +1279,7 @@ int ANSLPR_MultiGPU_StressTest() {
     //  CUDA cleanup (cuArrayDestroy, cuMemFree) while inference is running.
     //  This is the exact scenario that triggers the nvcuda64 SRW lock deadlock.
     // =========================================================================
-    std::atomic<bool> chaosEnabled{true};
+    std::atomic<bool> chaosEnabled{false};  // Disabled for VRAM stability long test
     std::thread chaosThread([&]() {
         std::mt19937 rng(std::random_device{}());