App Separation & Production Hardening

Highlights

• Two standalone Streamlit apps — Split the monolithic app.py into apps/embed_explore/ (embed & cluster your own images) and apps/precalculated/ (explore precomputed embeddings), each with dedicated entry points.
• Shared module architecture — Common code extracted into shared/{components,services,utils,lib}/ to keep both apps DRY.
• GPU-to-CPU fallback — Clustering auto-falls back through cuML → FAISS → sklearn on OOM or CUDA errors.
• cuML UMAP stability — UMAP runs in an isolated subprocess with L2-normalized embeddings to prevent SIGFPE crashes.
• Lazy-loaded heavy libraries — Deferred imports of torch, cuML, and FAISS cut startup time drastically
• Improved visualization — Zoom/pan support and heatmap option for clustering charts; chart interactions no longer trigger full page reruns.
• Dynamic metadata filtering — Precalculated app auto-generates filters from parquet schema.

Documentation

• Added BACKEND_PIPELINE.md, DATA_FORMAT.md, and .github/copilot-instructions.md for testing & code review
• Updated README with two-app workflow and simplified install instructions

Dependencies

• Relaxed numpy cap from <=2.2.0 to <2.3 (numba compatibility)
• Separate CLI entry points via pyproject.toml

Initial Combined App

This release contains the initial development of the image embedding explorer. It contains a working monolithic Streamlit application that allows users to embed local images and explore embeddings in 2D space, and also supports exploring precalculated embeddings.