Fix VLM: use Llama-4-Scout via nscale (confirmed working with images), Kimi-K2.6 and Qwen3.5-9B via together as fallbacks

Fix VLM models: use actual vision models with correct providers (Qwen2.5-VL, Gemma-4 via novita, Kimi via fireworks-ai)

Add Agentic Refinement tab: iterative image→pattern→3D→projection→compare→refine loop

Pass actual 2D pattern pieces to 3D builder — 3D view now matches 2D patterns

Fix Gradio 6.0 compatibility: move css/theme from Blocks() to launch()

Add chat editing + 3D preview + updated models

Upgrade to Qwen3.5-9B, Gemma 4, and Kimi-K2.5 (replace outdated Qwen2.5-VL)\n\nVerified working via Together AI provider:\n- Qwen/Qwen3.5-9B (primary VLM)\n- google/gemma-4-31B-it (Gemma 4)\n- moonshotai/Kimi-K2.5 (fallback)\n\nHandles reasoning+content response format from newer models."

Add main Gradio app with VLM integration