Simplify: Use ResNet-1D only (93.4% accuracy) - remove TCN and stacking

Fix: Use same text prompt 'Image' as training instead of 'Describe'

Fix: Correct meta-learner input format to match training (use real_prob not ai_prob)

Fix: Swap class interpretation - use class 0 as AI probability

Fix: Use aspect-ratio preserving resize to match training preprocessing

Enable API access with queue

Add torchvision dependency

Initial deployment with trained models