I have not looked into the literature. Is there a valid use case for y_true < 0?

There was a bit of discussion about this in the original issue #17639. I'm largely basing this off of Wang et. al. 2013 which has ~250 citations (details) and affirms the expectation that NDCG is contained within $[0,1]$. More directly, this survey paper on information-retrieval methods also uses relevance values (y_true) no less than 0.

Indeed, with the idea of relevance in document retrieval where NDCG was first formulated, it makes more sense for an additional document to offer no information and have no relevance, than to remove information or relevance from the overall query. From what I've read so far, including this additional survey paper talking about IR methods overall, cumulative gain (CG) is assumed to be monotone non-decreasing, so I don't think that y_true<0 is a valid use case, at the very least not for what it was designed to do.