VAAS: Vision-Attention Anomaly Scoring

This release extends VAAS with cross-attention fusion and resolves critical scoring inconsistencies, while maintaining a stable and reproducible inference interface. Highlights: Cross-attention fusion between global (Fx) and patch-level (Px) representations Corrected reference statistics computation (mu_ref, sigma_ref) for stable S_F scoring Resolved anomaly score collapse observed in earlier v2 builds Consistent behaviour across Hugging Face models and local checkpoints Release of v2 model variants (base, medium, large) on DF2023 Verified inference parity through integration tests Updated inference pipeline for improved reliability and usability Documentation and Colab examples aligned with the current API This version strengthens the coupling between global attention and local anomaly reasoning, enabling more reliable and interpretable anomaly scoring across different manipulation types.