fix: membenarkan masalah seed operasi-operasi kecil

refactor: replace chronological train/val split with optional holdout fraction

feat(serving): add configurable retraining data splits

refactor: extract training logic into dedicated training loop module