Sorry, I misspoke. Transformation is what makes the LLM itself legal -- its training data is sufficiently transformed into weights.
And so, a work being sufficiently transformative is one way in which copyright no longer applies, but that's not the case here specifically. The specific case here is essentially just a clean-room reimplementation (though technically less "clean", but still presumably the same legally). But the end result is still a completely different expression of underlying non-copyrightable ideas.
And in both cases, it doesn't matter what the original license was. If a resulting work is sufficiently transformative or a reimplementation, copyright no longer applies, so the license no longer applies.
That's interesting, but it misses my point:
The library's test suite and interfaces were apparently used directly, not transformed. If either of those are considered part of the library's source code, as the license's wording seems to suggest, then I think output from their use could be considered a work based on the library as defined in the license.