PackedQuadCell.getTokenBytesNoLeaf and PackedQuadCell.getTokenBytesWithLeaf have a couple issues that were exposed in
LUCENE-7845. One is that the WithLeaf version actually modifies the state (long term) which it shouldn't. Another is that if the BytesRef is null, it can compute a different result than if it's not null, which is definitely wrong.
It appears that these problems are only revealing themselves in practice with the
LUCENE-7845 optimization since there has been no PackedQuad test failure before, and RandomSpatialOpFuzzyPrefixTreeTest really hammers the prefix tree.