#3655 [GUI] GTX 1080 setup marked as unsupported by default (v0.2.1 / v0.2.2) and settings get reset

本文

### Area
GUI

### Summary
Tested across both **v0.2.1** and **v0.2.2**: the GUI incorrectly flags GTX 1080 GPUs (Pascal architecture) as unsupported/unrunnable and sets them to `Off` by default (showing *“No supported graphics card reported”* or *“cannot run Strata”*). 

However, when manually overriding the roles to `Main` and `Helper`, the underlying backend runs perfectly and shares the model across both GPUs (~23 tok/s, ~6.8 GB VRAM on each GPU). 

For context, my system uses the integrated GPU (iGPU) from the AMD Ryzen 7 5700G for the OS display[cite: 4], leaving both GTX 1080 cards entirely dedicated to AI inference.

### Additional Issues
1. **Lack of persistence / Reset:** Manual multi-GPU role assignments are not properly sticky or get overwritten, and running *“Tune this model”* completely overrides user choices back to a single GPU[cite: 6].
2. **JSON restrictions:** Trying to manually enforce splits or orders via configuration JSON triggers schema errors (`Unknown native setting` / schema validation blocks).

### Environment
- **Strata Versions Tested:** v0.2.1 and v0.2.2
- **OS:** Windows 11
- **Hardware:** 
  - System Display / iGPU: AMD Ryzen 7 5700G with Radeon Graphics[cite: 4]
  - AI Inference: 2x NVIDIA GeForce GTX 1080 (8GB each) + 56GB DDR4 RAM[cite: 1]
- **Model:** Qwen3.8-Flash-Next Q2_0[cite: 1]

### Steps to Reproduce
1. Clean install of Strata (v0.2.1 or v0.2.2) on a multi-GPU setup.
2. Observe that the GUI automatically disables or flags the GTX 1080 cards as unusable[cite: 10, 13].
3. Manually set GPU 1 to `Main` and GPU 2 to `Helper`[cite: 2]. The model loads and runs successfully across both cards[cite: 5].
4. Restart the app or trigger tools like "Tune", which forces a reset of the hardware layout.

### Expected Behavior
- Fully functional CUDA-compatible GPUs (like the GTX 1080) should not be blocked or disabled by default just because they lack newer architectural features, especially when an iGPU handles the display[cite: 4, 10, 13].
- Manual multi-GPU selections should remain persistent across restarts and not be overridden by auto-detection heuristics.

<img width="1357" height="894" alt="Image" src="https://github.com/user-attachments/assets/70e96916-9789-4b1e-a1d6-942f1f92c968" />