I'd imagine part of it is that there are xpath parsers on github already, so it's not entirely greenfield for the models.

Partially, but I think it's moreso the W3C spec being so thorough and the wealth of samples online. I found that during code review both Codex and Claude would refer back to the specs for certainty on expected outcomes. Their understanding of how to deal with unusual edge cases was also impressive, so it seems they have a lot of baked in knowledge & training of how XPath works and they draw from that.