Fix CUDA fallback: gracefully fall back to CPU when CUDA libs missing

- transcribe: catch model load failures on CUDA and retry with CPU
- hardware detect: test CUDA runtime actually works (torch.zeros on cuda)
  before recommending GPU, since CPU-only builds detect CUDA via driver
  but lack cublas/cuDNN libraries

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

python/voice_to_notes/hardware/detect.py

@@ -105,14 +105,23 @@ def detect_hardware() -> HardwareInfo:
     # RAM info (cross-platform)
     info.ram_mb = _detect_ram_mb()
 
-    # CUDA detection
+    # CUDA detection — verify runtime libraries actually work, not just torch detection
     try:
         import torch
 
         if torch.cuda.is_available():
-            info.has_cuda = True
-            info.cuda_device_name = torch.cuda.get_device_name(0)
-            info.vram_mb = torch.cuda.get_device_properties(0).total_mem // (1024 * 1024)
+            # Test that CUDA runtime libraries are actually loadable
+            try:
+                torch.zeros(1, device="cuda")
+                info.has_cuda = True
+                info.cuda_device_name = torch.cuda.get_device_name(0)
+                info.vram_mb = torch.cuda.get_device_properties(0).total_mem // (1024 * 1024)
+            except Exception as e:
+                print(
+                    f"[sidecar] CUDA detected but runtime unavailable: {e}. Using CPU.",
+                    file=sys.stderr,
+                    flush=True,
+                )
     except ImportError:
         print("[sidecar] torch not available, GPU detection skipped", file=sys.stderr, flush=True)