SKLearn train_test_split jumbles indexes - unreproducable with dummy data

Question

I have a setup where I read a csv into a dataframe, add a calculated column, then do a train_test_split. A dummy solution would be:

import pandas as pd
import random
asd = random.sample(range(1, 1456165166), 500)

index = list(range(500))

data = {
  "calories": asd,
  "lol": asd,
  "ix": index
}
df = pd.DataFrame(data)
df = df.set_index("ix")
df['target_cat'] = np.where(df['lol'] > 154683526, 0, 1)
X = df.loc[:,df.columns != "lol"]
y = df['lol']
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3)

This works perfectly, and it keeps the indices in X_train and y_train consistent, e.g.:

This would be the expected behavior. Now if I do the same with my own data read from a csv, all the indices get jumbled:

train = pd.read_csv(all_files[2], delimiter=",", index_col=None, header=0, encoding="UTF-8")
train = train.set_index("id")
train['target_cat'] = np.where(train['target_reg'] == 0, 0, 1)
X = train.loc[:,train.columns != "target_cat"]
y = train['target_cat']
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3)

When I do this, the indices won't line up, seem to be off by 1:

How on Earth can this happen? I've been sitting here baffled and have no idea what I'm missing...