What does this PR do?

Fixes a TypeError raised by HDBSCAN when cluster_selection_epsilon is set and the input data contains tied distances.

Fixes #33219

Root cause

In traverse_upwards (sklearn/cluster/_hdbscan/_tree.pyx), two lookups on the condensed cluster tree use boolean indexing:

parent     = cluster_tree[cluster_tree['child'] == leaf]['parent']
parent_eps = 1 / cluster_tree[cluster_tree['child'] == parent]['value']

When tied distances exist in the data, the condensed tree can contain multiple rows with the same child id. The boolean-index lookups therefore return a shape-(1,) array instead of a scalar. NumPy 2.4+ removed the implicit array-to-scalar conversion that had silently papered over this, causing:

TypeError: only 0-dimensional arrays can be converted to Python scalars

Fix

Add explicit [0] indexing on both lookups. All duplicate rows share the same parent and lambda value, so taking the first element is correct:

-    parent = cluster_tree[cluster_tree['child'] == leaf]['parent']
+    parent = cluster_tree[cluster_tree['child'] == leaf]['parent'][0]
 ...
-    parent_eps = 1 / cluster_tree[cluster_tree['child'] == parent]['value']
+    parent_eps = 1 / cluster_tree[cluster_tree['child'] == parent]['value'][0]

This matches the behaviour of the original hdbscan package from which scikit-learn's implementation was ported.

Tests

Added test_hdbscan_cluster_selection_epsilon_tied_distances in sklearn/cluster/tests/test_hdbscan.py using the exact precomputed distance matrix from the issue report. The test verifies:

• No TypeError is raised
• All non-outlier points are assigned to a single cluster

Full HDBSCAN test suite: 149 passed, 0 failed.

Note on triage

The betatim scikit-learn member confirmed the fix approach in the issue thread and provided the exact diff, so this PR implements that confirmed fix with a tested, clean implementation.