A reasonable critique, in part, but you are missing some important stuff:
SQL:2003 added Window Functions, standardized sequences, and columns with auto-generated values
SQL:2006 okay, this one sounds like it was mostly about XML, but a lot of the functionality that went into dealing with structured content is what made the JSON support workable.
SQL:2008 this gave us the MERGE version of upsert, instead of triggers (which apply to views, and are a powerful way to deal with backwards compatibility)
SQL:2016 LISTAGG (super powerful), Row Pattern Recognition (a lot more intuitive that comparable uses of WINDOW functions), standardized date/time parsing, and polymorphic table functions (which you can kinda see in, for example, BigQuery's legacy table wildcard functions)
This is pretty much what I mean by "willy-nilly". Little concern seems to be paid to how well it integrates into some holistic information access theory or paradigm. I would expect some adherence to relational algebra, and if that isn't sufficiently powerful, I would like a transition to some more powerful theory rather than welding on the new hotness. I'm not saying I could do it better, mind you.
SQL:2003 added XML functions
SQL:2006 more XML!
SQL:2008 XQuery!
SQL:2016 JSON!
My guess it's not so much "willy-nilly" but more a sign of "what was popular."