Feature stores and reproducible training pipelines

I’m verifying the current Feast release and APIs, then I’ll deliver the complete Markdown article with a runnable local pipeline and the train/serve skew case study.