BUG REPORT: Inverse Boolean Logic in "Limit Model Offloading to GPU Dedicated Memory" Toggle affecting MoE Models

Nobody has claimed this yet.

### System Environment:

- **LM Studio Version:** 0.4.17
- **OS:** Windows 10
- **Hardware:** AMD Ryzen 7 5700X3D | NVIDIA RTX 5060 Ti (16GB VRAM) | 32GB RAM
- **Model Tested:** Qwen 3.6 35B A3B (GGUF) with large context 64K

### Description of the Bug & Visual Evidence:

There is an inverse boolean logic bug with the **"Limit Model Offloading to GPU Dedicated Memory"** toggle when running `Qwen 3.6 35B A3B`. The behavior of the switch is completely inverted compared to its visual label:

- **When ON (Toggle Enabled):** VRAM caps at ~13.5 GB, offloading processing to the CPU (causing CPU usage to spike). Performance drops to **~23 tok/s**.

- **When OFF (Toggle Disabled):** The backend correctly allocates maximum VRAM (~15.2 GB), keeping shared memory at 0 GB, and performance flies at **+72 tok/s** with low CPU usage.

Start by reproducing the report on Windows 10 with LM Studio 0.4.17, the listed RTX 5060 Ti hardware, and Qwen 3.6 35B A3B at large context. Compare CPU usage, VRAM allocation, shared memory, and token rate with the “Limit Model Offloading to GPU Dedicated Memory” toggle on and off; done means the label and behavior agree without the reported performance regression.

Written by the indexing model from the issue text.