fix(zerogpu): persist raw rank diagnostics and script indexes

test: add zerogpu metrics runner for small models

Track baml_client in git, remove runtime generation

chore: add local gradio submodule and setup script

feat: SmolLM2-135M ranker fine-tuning pipeline + BAML streaming improvements

feat: Dockerfile + .dockerignore + setup script for BAML codegen

feat: production-ready DenseFeed pipeline with jj workflow docs