Remove torch (~800MB) from build — use pickle for .pt files — build time: 10min→2min

Restore profiles in Explorer: disp_profiles.json with 36k sparklines

Fix: export PETIMOT_ROOT from data_loader.py

data_loader: auto-download ground_truth from HF Dataset at startup

Add ensure_ground_truth() - auto-extract ground_truth.zip on HF Space startup

Fix: missing embedding_viewer, NaN in correlation matrix, upload all components

Add disp_profile sampling for sparklines

Fix: handle empty index.json, fallback to zip scan

Fix: auto-download predictions.zip from HF if not found or LFS pointer

Fix: LFS pointer detection + debug logging for predictions

Add LFS pointer detection + debug logging

Fix: robust zip reading + handle empty DataFrame

Update data_loader to read directly from zip

Add Streamlit explorer app with Docker deployment