[b84330808]

partition panics when the lanes along axis have a negative stride:

use ndarray::prelude::*;

let mut a = array![[7, 1, 5, 3], [2, 6, 0, 4]];
a.invert_axis(Axis(1));
a.partition(1, Axis(1)); // panics

let b = array![3, 1, 4, 1, 5, 9, 2, 6];
b.slice(s![..;-1]).partition(3, Axis(0)); // panics too

called `Option::unwrap()` on a `None` value

partition copies the array with to_owned(), which keeps the negative stride, and then takes the slice fast path if the first lane is_contiguous(). A lane with stride -1 counts as contiguous, but as_slice_mut() only returns a slice for standard layout, so the unwrap() in the fast path fails.

This checks as_slice_mut().is_some() instead, which is exactly what the fast path needs. Lanes with a negative stride now go through the existing copy path, which already handles any stride.

Added test_partition_negative_stride covering an inverted axis and a reversed view; it panics without the change. cargo test -p ndarray -F approx,serde,rayon, cargo clippy -F approx,serde,rayon and cargo +nightly fmt --all --check pass locally.

I used Claude Code to write this change.

[by-ae]

Checked this out at c4aba74 and ran the partition tests in the lib (cargo test -p ndarray --lib test_partition): all 8 pass, including the new test_partition_negative_stride.

I also ran a randomized probe outside the test suite: 3000 random 1-3D i32 arrays (C and F order), with each axis randomly inverted or sliced with step -2 or 2, partitioning along a random axis at a random kth. For every lane it checked that the result is a permutation of the input lane, that p[kth] equals the kth element of the sorted lane, and that everything before or after it is <= or >= respectively. On master the same probe panics at the as_slice_mut().unwrap() inside the fast path (impl_methods.rs:3324). On this branch all 3000 cases pass.

Checking only the first lane is fine, since every lane along axis has the same stride, so if as_slice_mut() works for the first lane it works for all of them.