Your test is more-or-less contrived to fail given the tradeoff to avoid repeated memory allocations but that doesn't say much about the actual usage in uuid generation since it's not exported for general purpose use.

Presumably they had some hot path somewhere where rng() is called in a loop and this optimization made sense with awareness that it could be misused as in your example breaking the contract ensuring randomness, which (hopefully) they're not actually doing anywhere.

Unless I'm missing something replacing the package over this with a less vetted implementation seems excessive and possibly even counterproductive.