Two nodes, one directed edge: target gets more rank.

A dangling target node (only incoming edges) should still get rank.

A single node with a self-loop should get all the rank.

Should skip edges where source == target.

All nodes point to hub — hub gets most rank.

Hub points to all — leaves get more rank than hub.

Test that processing stops early when an error occurs and fallback is None.

Test async_map with string inputs.

Test text embedding with each embedder.

Test that auto modality correctly detects text.

Three nodes in a cycle: equal rank of 1/3 each.

Very tight tolerance should still produce valid results.

Explicit n includes trailing nodes that have no edges.

Two nodes in a cycle: equal rank of 0.5 each.

Test that 'sentence_transformers' (underscore) is accepted and normalized. We can't do a full integration test without a real model, so we verify

Unweighted (2-tuple) edges should work.

Very large weights should not overflow.

Very small weights should not cause division-by-zero.

Higher weight edge should direct more rank to its target.

Zero-weight edges should not cause errors.

A simple DataFrame with text data.

Run UMAP benchmarks.