OsaurusAI/Qwen3.5-35B-A3B-JANG_4K

capabilities: stamp reasoning/tool parser + cache_type for vmlx CapabilityDetector Tier-1

Add MMLU benchmarks (200q, thinking OFF, greedy)

Update usage to Osaurus branding

Qwen3.5-35B-A3B-JANG_4K JANG mixed-precision quantization for MLX

initial commit