Karpathy-style recipe and common-pitfalls checklist for training neural networks from scratch. Use when designing, debugging, or reviewing a custom training pipeline (data loading, loss, optimizer, regularization) — especially when training silently fails to converge or underperforms with no errors thrown.
