[yalinli2]

Model.load_samples(samples) sorts samples with a nearest-neighbour walk over the coupled-parameter columns. Two inputs made that sort crash. sort=False avoided both.

1. Samples that share the same coupled-parameter values

Raised AssertionError: issue in sorting algorithm (or KeyError) when several samples have identical values for all coupled parameters and differ only in isolated parameters. Morris and Sobol designs produce many such samples. I saw it with Morris N >= 20 and Sobol N = 256, on models with 6 to 11 parameters of which 1 to 3 are coupled.

Cause: the walk assumed argsort puts each row itself at position 0 of its neighbour list, and scanned from position 1. With ties at distance 0, another duplicate can land at position 0 and is never considered. If it is the last remaining sample, the loop makes no progress and the assertion fires. The initial step (nearest_arr[0, 1]) had the same flaw and could pick row 0 itself, raising KeyError.

Fix:

• Scan neighbours from index 0 instead of 1. The current sample is already removed from remaining, so this is safe.
• Start the walk at sample 0 and let the loop pick its nearest neighbour, replacing the special first step.

2. A coupled parameter that is constant across all samples

Raised FloatingPointError (when numpy errors are set to raise; otherwise the distances became NaN). The column has zero range, so normalizing (samples - min) / (max - min) divided by zero.

Fix: replace zero ranges with 1 before normalizing, so a constant column normalizes to zeros.

The result is still a greedy nearest-neighbour path.

[yoelcortes]

Thanks for the improvements. Reading the code, it looks straight forward. But just to make sure, I also checked the outputs with the old version and they match.