v3 model retrain on verdict_overrides corpus

After 2-3 weeks of customers marking spam/ham via the UI and digest links, we'll have a labelled corpus from real production. Retrain the LightGBM model on that, plus the bodies.jsonl training log.