I'm not too worried about this, and I think Gumroad's concern is likely overblown. I can't tell from their comment whether they actually experienced AI being bad at HTMX, or if they transitioned to talking about other resources.
LLMs are often wildly good at being universal translators. So if they pick up general patterns and concepts in popular frameworks, and enough syntax of more niche frameworks, IME, they do a pretty great job of generating good niche framework code.
Similar to how they can talk like a pirate about things things pirates never said in their training data.