Introduced Tuna-2, a native unified multimodal model using pixel embeddings

Documented failure to generate coherent world maps, becoming a benchmark for spatial reasoning weaknesses