"""Shared dataset splitting so every script uses the same held-out events."""
import numpy as np


def split_by_event(df, test_frac=0.15, seed=42):
    """Return (train_mask, test_mask, test_events) grouped by whole events."""
    rng = np.random.default_rng(seed)
    events = np.array(df["event"].unique())
    rng.shuffle(events)
    n_test = max(1, int(len(events) * test_frac))
    test_events = set(events[:n_test].tolist())
    mask = df["event"].isin(test_events).to_numpy()
    return ~mask, mask, test_events