docs: narrower tables to fix HF render overflow; clearer anomaly-class breakdown

docs: correct training recipe — Adam not AdamW, per-model LR/batch table, ViT 5e-4/5e-5, no scheduler in two-stage, AMP details

docs: add model card with v2 metrics, training recipe, citation

Update ensemble weights (v2-2026-05-07): 10+10 epoch two-stage runs across 5 backbones; Celeb-DF ensemble AUC 0.8851, gap 0.1149

Add model card with performance tables and usage

Add metadata.json

Add r3d18_raft_best.pth

Add vit_base_patch16_224_best.pth

Add r3d18_best.pth

Add efficientnet_b4_best.pth

Add resnet18_best.pth

initial commit