lineardiff returns NaNs when step_size is the same as window_size

[alexzerg]

Module lineardiff, the same value for step_size and window_size. It looked like a normal combination to try but some parts of the result become NaN.
Current joss2.0, python 3.13:

import numpy as np
from pynumdiff.linear_model import lineardiff

t = np.linspace(0, 2*np.pi, 41)
dt = t[1] - t[0]
x = np.sin(t)

for step in [10, 11, 12]:
 x_hat, dxdt_hat = lineardiff(
 x,
 dt,
 order=2,
 gamma=0.01,
 window_size=11,
 step_size=step,
 solver=“CLARABEL”,
 )

 print(
 step,
 np.sum(~np.isfinite(x_hat)),
 np.sum(~np.isfinite(dxdt_hat)),
 )

—>

10 0 0
11 4 6
12 6 12

There are also invalid value encountered in divide warnings. Some samples aren't covered by any window when the stride gets larger so their weight_sum stays zero and slide_function divides by it at the end.
I didn't see a restriction for step_size in the docstring. Values which leave samples uncovered should be rejected before running instead of returning partially NaN arrays. What do you think Pavel?

Review reference: openjournals/joss-reviews#11172

[pavelkomarov]

If the stride is wider than the window length, then you're leaving some holes as you go. Claude also tells me it's possible the last few elements in an array can go uncovered by any window. Fixing both.

[pavelkomarov]

26f0a53