fix(models): validate inputs upfront and fix pickling of the Lightning module

[ChrisW09]

Fixes #424 (the three items with a clear single fix; the remaining items in that issue — the MLflow sqlite directory and the InferenceModel double-load — are left for separate PRs).

Problems

1. __getstate__ nulls the wrong key. state["task_model"] = None — but the attribute is _task_model, so pickling kept the full Lightning module and merely added a spurious task_model key. The docstring in _mixins/serialization.py explicitly documents the opposite intent.
2. fit() crashes on plain-list X. The fit.started event evaluates X.shape[1] if hasattr(X, "shape") else len(X.columns) eagerly, before ensure_dataframe runs in _build_model — so list input dies with AttributeError: 'list' object has no attribute 'columns' even though the converter handles lists fine.
3. All-zero sample_weight fails cryptically. It reached torch.multinomial and raised a RuntimeError mid-training. This is the one failing check in the repo's own suite today: test_sklearn_contract.py::...::check_all_zero_sample_weights_error, which requires a ValueError matching .*weight.*zero.*.

Fix

Null _task_model; fall back to len(X[0]) for list input in the event payload; validate sample_weight in _resolve_loss_and_sampler (non-negative, at least one non-zero) with actionable messages.

Tests

New tests/test_api_robustness.py covers all three, plus a positive case confirming valid non-uniform weights still work. tests/test_sklearn_contract.py now passes fully green (previously 1 failure), and tests/test_class_imbalance.py is unaffected.

🤖 Generated with Claude Code