[TheNetherWatcher]

resample_sig returns only NaN values if the signal starts with NaN:

import numpy as np
from wfdb.processing import resample_sig

resample_sig(np.array([np.nan, 1.0, 2.0, 3.0, 4.0, 5.0]), fs=2, fs_target=4)[0]
# array([nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan])

NaN values are filled with pd.Series.interpolate() before resampling (added for #249), but by default it only fills forward. Trailing NaN values take the last valid value, while leading ones stay NaN, and the FFT resample then spreads them over the whole output. This also affects resample_singlechan and resample_multichan, for example on a record whose first samples are invalid.

The fix passes limit_direction="both", so leading NaN values take the first valid value. Inner and trailing NaN values are filled as before.

Tests: a signal with leading NaN values gives the same result as the signal with them filled (this fails on main), and a test for inner and trailing NaN values checks that they are still handled as before. pytest -m "not network" passes (98 tests), and the docs build passes.

[tompollard]

Thanks, this looks good! I checked that the leading-NaN test fails on main, that an all-NaN signal behaves as before, and that record 100 with its first 50 samples set to NaN now resamples to finite values.

One note for later, not for this PR: since the filled values are resampled like real data, the output has no NaN where the input was invalid. It might be worth marking those regions as NaN in the output at some point, but that's a separate change.