[antonsoo]

Integer samples retain their dtype through extract_features, so energy calculations can overflow before returning a floating result. For example:

import numpy as np
import pandas as pd
from tsfresh import extract_features

frame = pd.DataFrame({
    "id": [0] * 4,
    "sample": np.array([20000, 30000, -20000, -30000], dtype=np.int16),
})
features = extract_features(
    frame, column_id="id", n_jobs=0, disable_progressbar=True,
    default_fc_parameters={"abs_energy": None, "root_mean_square": None},
)
print(features.iloc[0].to_dict())
# Before: {'sample__abs_energy': -9728.0, 'sample__root_mean_square': nan}
# After:  {'sample__abs_energy': 2600000000.0,
#          'sample__root_mean_square': 25495.097567963923}

Promote integer and Boolean inputs to float64 before squaring or taking the dot product in abs_energy, root_mean_square and energy_ratio_by_chunks. Chunk ratios have the same problem: the numerator uses a floating exponent but the denominator currently squares in the integer dtype. Floating input dtypes and unrelated calculators are preserved.

Regression checks cover signed and unsigned integers from 8 to 64 bits, Boolean arrays/Series, uneven chunk partitions, and native extraction. The full suite passes (321 passed, 14 skipped). Changed-file hooks pass; all-file hooks reproduce one existing README whitespace violation on unchanged main.

Additional checks load speech and music into 8/16/24/32-bit PCM WAVs and extract 192 windows through serial and multiprocessing paths. Independent Python-integer sums of squares identify 926 wrong feature values before the change and none after it. All 384 floating-input windows and six unrelated features across 576 rows remain byte-identical. This checks arithmetic and extraction behavior, not classification quality. Integer inputs require a float64 work array; full-width integer energy remains a floating approximation.