✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: 02cadbf. Link to the linter CI: here}

So the hdbscan one seems recent since it was not in my previous commit, this seems like a numpy change that somehow changes the HDBSCan labels, to be investigated:

numpy dev

n_clusters=3
labels={np.int64(0), np.int64(1), np.int64(2), np.int64(-1)}
centers=[(0.0, 0.0), (3.0, 3.0)]
hdb.centroids_=array([[ 3.00974728,  2.99780391],
       [-0.05690994,  0.00368199],
       [ 1.0671027 , -0.61185795]])

numpy release

n_clusters=2
labels={0, 1}
centers=[(0.0, 0.0), (3.0, 3.0)]
hdb.centroids_=array([[-0.04358032, -0.0044484 ],
       [ 3.00974728,  2.99780391]])

So It looks like np.argsort has slightly changed in numpy dev (likely the optimization in numpy/numpy#25610). This changes the centroids in HDBSCAN. Adding more samples makes the test more robust.

A snippet that shows the numpy dev change:

import hashlib
import numpy as np
hashlib.sha256(np.argsort(np.random.default_rng(0).integers(100, size=10_000)).tobytes()).hexdigest()

numpy 1.26.3

97d7de848fa6a7df48d2d0ec7608f0a772c64f8a4d9e33c2c754c5be6030beaf

numpy dev

dcc8d89375bc7463d539315c342a6a9b40493451116a531b1e97949cc6f44ce8

np.argsort use quicksort by default which is not stable and we use np.argsort it with default parameter in HDBSCAN so I guess there was no guarantee. The np.argsort optimization in numpy-dev is changing the output enough that HDBScan finds 3 centroids instead of 2.