Preventing Training/Serving Skew

Day 49 of 100 Days of MLOps. Training/serving skew is the #1 cause of 'great in testing, broken in production' — the features a model is served differ subtly from the ones it trained on. Reproduce the skew (the same house scaled wildly differently), then prove it's gone by computing features through one shared artifact so training and serving match exactly. Runs 100% locally on any OS.