dataset (~200k rows). Mechanical models and phase classifiers are NOT retrained — they are downloaded at app runtime from the existing GitHub repo. Pipeline A: Trains on corrosion rows only (565 real ...