Scaffold a clean PyTorch training loop with eval and early stopping

You are a senior ML engineer who writes clean, reproducible PyTorch code.

Task: [CLASSIFICATION / REGRESSION / SEGMENTATION — and the data, e.g. '30k labeled tabular rows, 40 features, 5 imbalanced classes'].
Runtime: [single GPU / CPU / multi-GPU DDP]. PyTorch [2.x].

Produce a single runnable train.py with:
1. A config dataclass for every hyperparameter (lr, batch size, epochs, weight decay, seed, device) with sane defaults and a fixed seed for reproducibility.
2. A model module matching the task, with configurable depth/width and dropout — small and correct, not novel.
3. Weighted loss + sampler/pos_weight handling if the data is imbalanced.
4. train() and evaluate() functions: running loss, accuracy/F1 as appropriate, tqdm progress, and zero gradient leaks in eval.
5. Simple early stopping on the validation metric with patience [N].
6. Checkpoint best-val weights to disk; log to CSV or behind a Weights & Biases flag.
7. An argparse CLI so any config value can be overridden.

Guardrail: this is an engineering scaffold, not research advice. Do not claim SOTA results or invent metrics. Say plainly where I must supply real data and labels.

Output the full script in one fenced block, then a 4-item runbook with exact launch commands.

Success signal: the output is good only if it runs with one command, seeds are fixed, eval has no gradient leaks, and early stopping triggers on the validation metric.