This sounds like a bug in postgres (or a low-hanging-fruit optimization); adding the hacky offset 0 to bypass the inlining may help work around the bug now but

This sounds like a bug in postgres (or a low-hanging-fruit optimization); adding the hacky offset 0 to bypass the inlining may help work around the bug now but may hinder future optimizations. My experience teaches me to avoid tricking an optimizer down a path. Write the query in the most easy-to-understand way to avoid massive future pain. I would also avoid json/xml 'support', it's a recipe for unintended consequences.