[antonsoo]

The per-vector loops slice x.shape with the raw axis. With a negative axis this can include the time dimension among the vector indices, leading to an indexing error or results along the wrong dimension. For example:

import numpy as np
from pynumdiff import splinediff

x = np.arange(18.0).reshape(2, 9)
x_hat, dxdt_hat = splinediff(x, .1, axis=-1)
print(x_hat.shape, dxdt_hat.shape)

Before, this raises IndexError: too many indices for array. After, both shapes are (2, 9), with the same values as axis=1.

Normalize the axis before building vector index tuples in splinediff, polydiff, rtsdiff, robustdiff and tvrdiff. Indexing range(x.ndim) resolves valid negative axes and rejects out-of-range axes. The numerical algorithms and positive-axis paths stay the same.

Validation: all 225 tests pass, including equivalent-axis and input-preservation regressions. Pylint was run against both sources; existing diagnostics remain. In 108 recorded KITTI velocity/gyro workflows with scalar and irregular timing, 30 exceptions and six incorrect results become zero. All 54 positive-axis control hashes remain exact. Separate cubic-spline and dense Gaussian trajectory references agree within 3.4e-9. The existing convex-solver accuracy warning is also present on the unmodified base.