cvikl/religious-artwork-analysis

Warm numba kernels at startup so first reprojection is fast

Pre-reduce wide feature groups to 50 PCA dims for fast reprojection

Instant return to default layout for all-ones weights

Warm-start t-SNE reprojection (~4x faster weight changes)

App: data/ paths

Space: code only; data fetched from HF dataset repo at startup

First commit.