Return NaN for non-finite input instead of a made-up entropy

[Par-python]

Problem

A single NaN or ±inf in the input produced a plausible-looking entropy instead of NaN:

x = np.random.default_rng(0).standard_normal(64); x[10] = np.nan
permutation.compute(x)                    # 2.5336 (argsort ranks NaN last, counted as a pattern)
spectral.compute(x)                       # -0.0   (NaN spectrum; the psd > 0 filter keeps nothing)
permutation.compute(np.full(64, np.nan))  # -0.0

rolling inherited this, so one gap in a logged metric silently corrupted every window that contained it. It was found while building training-early-warning, where spectral entropy dropped to exactly 0 in a window that ran past a NaN loss.

Fix

• _core.nan_on_non_finite wraps every measure's _kernel: compute, normalized, rolling and delta return NaN when the input or window holds NaN or ±inf.
• rolling is NaN for exactly the windows containing the gap; every other window is unchanged.
• The same guard covers the paths that bypass the kernels: multiscale.compute, transfer, spectral.normalized, and divergence.kl / js.
• Zero entropy is returned as 0.0, never -0.0.
• README documents the missing-value behavior and that rolling is causal. CHANGELOG marks the change as one that changes results.

EntropyFeatures is unchanged: scikit-learn's check_array already rejects NaN input with a clear ValueError, as every sklearn estimator does.

Tests

tests/test_missing_values.py: NaN, inf and -inf for every single-series measure, all-NaN input, rolling windows NaN only around the gap and identical elsewhere, delta, multiscale, transfer (both estimators, gap in x or y, rolling), divergence, and positive zero. The full suite passes locally (248 tests).

Local ruff is 0.15.15 while CI pins 0.16.7, so it's worth watching the lint job.

🤖 Generated with Claude Code