Few methods silently accept invalid input.

[sidxz]

Nit picking here!, but it's a small robustness thing.
Some methods take dt_or_t and some handle NaNs, it's easy to try either on a method that doesn't support it.
Right now those cases don't complain, they just return inf or NaN.

Few that I found:

Time array into a fixed-step method (since t[0] == 0 it ends up dividing by zero):

import numpy as np
from pynumdiff.polynomial_fit import savgoldiff
from pynumdiff.finite_difference import finitediff

dt = 0.01; t = np.arange(0, 3, dt)
x = np.sin(3*t) + 0.05*np.random.default_rng(0).standard_normal(len(t)); true_dxdt = 3*np.cos(3*t)

_, d_ok = savgoldiff(x, dt, degree=2, window_size=9, smoothing_win=9)
_, d_bad = savgoldiff(x, t, degree=2, window_size=9, smoothing_win=9) # t by mistake
print(np.abs(d_ok - true_dxdt).max(), np.abs(d_bad - true_dxdt).max()) # 1.51 inf
print(np.abs(finitediff(x, t)[1] - true_dxdt).max()) # inf

spectraldiff does refuse, but with operands could not be broadcast together with shapes (1000,) (300,), which doesn't really say what went wrong.

NaN into a method :

from pynumdiff.basis_fit import spectraldiff
xn = x.copy(); xn[100] = np.nan
print(np.isnan(spectraldiff(xn, dt, high_freq_cutoff=0.2)[1]).sum()) # 300 of 300

kerneldiff does something similar (the NaNs spread through the convolution).
savgoldiff refuses via scipy's own check, which is fine.

This is actually very well handled in waveletdiff.

ValueError: `dt` must be a scalar. The DWT requires uniformly sampled data. For variable step sizes, use rbfdiff or splinediff instead.

Again, this one is a non blocker for me, nevertheless, reporting this for the better!

Review reference: openjournals/joss-reviews#11172