[C95234]

transition=0 is silently ignored. The constructors and about a dozen methods test truthiness (if self.transition:, float(transition) if transition else transition), and 0 is falsy, so a series indexed from a negative time with the tipping point at time zero is treated as if no transition had been given: every indicator is computed over the whole series, past the transition.

Reproduction on current main:

import numpy as np, pandas as pd, ewstools
rng = np.random.default_rng(0)
s = pd.Series(rng.normal(size=100).cumsum() * 0.1, index=np.arange(-50, 50))

for tr in (0, None, 30):
    ts = ewstools.TimeSeries(s, transition=tr)
    ts.compute_var(rolling_window=0.25)
    print(tr, ts.ews["variance"].notna().sum())
# 0    76   <- same as no transition
# None 76
# 30   62

MultiTimeSeries behaves the same way: with transition=0, compute_corr runs to the last time point (49) as without a transition, while transition=30 stops at 30.

Fix: is not None instead of truthiness in the two constructors (float(transition) if transition is not None else None), in the 13 if self.transition: checks (detrend, compute_var, compute_std, compute_cv, compute_auto, compute_skew, compute_kurt, compute_entropy, compute_dfa, compute_spectrum, and detrend, compute_covar, compute_corr on MultiTimeSeries), and in apply_classifier_inc (not self.transition). Behaviour for transition=None and for non-zero transitions is unchanged; transition is still stored as a float.

Tests: tests/test_transition_zero.py (15 tests). 13 of them fail on main and pass with the fix; the other two pin the unchanged behaviour for None and for a positive transition. They cover compute_var, compute_std, compute_cv, compute_skew, compute_kurt, compute_auto, compute_dfa, detrend, compute_spectrum, and the MultiTimeSeries correlation, covariance and detrend. One test checks that a series with transition=0 gives the same indicators as the same data shifted by +50 with transition=50. Full suite: 46 passed, 1 skipped (the TensorFlow one).

Not covered by a test: compute_entropy and apply_classifier_inc (the latter needs a TensorFlow classifier). The change in both is the same one-line condition.

CHANGELOG.md has an entry under [Unreleased].

One thing left alone on purpose: float(transition) still rejects a pd.Timestamp; that is a separate question.