Fix model selector not showing loaded GPU model on CPU-detected hardware

The /models endpoint filtered models by hardware capability, hiding
GPU-only models (Devstral) when CUDA init failed (e.g. driver too old).
But the model was already loaded and running on CPU successfully.

Now always includes the currently loaded model in the available list,
regardless of hardware detection. If it's running, it's available.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

backend/model_service.py

@@ -1311,17 +1311,23 @@ async def list_models():
     if has_gpu and torch.cuda.is_available():
         available_vram = torch.cuda.get_device_properties(0).total_memory / (1024**3)  # GB
 
+    # Always include the currently loaded model (it's running, regardless of
+    # what the hardware check thinks — e.g. CUDA driver too old but CPU works)
+    current_model_id = manager.model_id if manager else None
+
     models = []
     for model_id, config in SUPPORTED_MODELS.items():
         model_min_device = config.get("min_device", "cpu")
 
-        # GPU backends can run all models
-        # CPU backends can only run CPU models
-        if device_type == "gpu" or model_min_device == "cpu":
+        # Include if: GPU backend, or CPU-compatible model, or currently loaded
+        if device_type == "gpu" or model_min_device == "cpu" or model_id == current_model_id:
             # Check VRAM requirements for GPU models
             is_available = True
             if has_gpu and available_vram > 0 and available_vram < config["min_vram_gb"]:
                 is_available = False
+            # Currently loaded model is always available
+            if model_id == current_model_id:
+                is_available = True
 
             models.append({
                 "id": model_id,