Could spectral initialization give unexpectedly large initial coordinates?

This might not be a bug: I haven't seen it cause a problem in practice, but I also haven't tried very hard to trigger it.

The spectral initialization involves rescaling the eigenvectors by maximum value, then multiplying by 10:
https://github.com/lmcinnes/umap/blob/d04006e6399f02a70bc831dd969a119e86048de4/umap/umap_.py#L945
(finally, some noise is added, but that's not germane to the current discussion)

I think the idea here is rescale the coordinates so that the largest value is 10. But if the coordinates can take negative values (which they can) and if the magntiude of the most negative coordinate is larger than that of the most positive coordinate (this bit I don't know about), the actual rescaling could result in the negative coordinates being much larger than 10.

I mention this because just such an unfortunate state of affairs has occurred in uwot, leading to horrible outliers in some embeddings (obviously, the fix is to use the maximum absolute value). I don't know much about the scipy implementation of the eigenvector routines that prevent negative values being larger than the positive ones, but I thought I would ask.

Start in umap/umap_.py at the spectral initialization code around line 945. Inspect how the scipy eigenvectors are rescaled and whether negative coordinates can exceed the intended bound. Reproduce or construct an asymmetric-coordinate case if possible, then verify the initialization remains bounded and add a regression test if the behavior is confirmed.