To pull this particular style of trick you require a schema definition that allows for one object to be expanded into a whole set of objects, and for the resulting data structure to be a tree rather than a simply a rooted directed graph.

I don't know YAML well but I believe if you tried this trick with something like alias nodes then you would end up with a lol9 node with ten separate connections to a single lol8 node with ten separate connections to a single lol7 node and so on. This would not produce the same problem in the parser, though might trigger problems in whatever processed the resulting graph.

That's definitely correct: Since you can produce cycles in YAML (it's deliberate choice), programs which don't check for graph cycles and blindly go about traversing a serialized graph are subject to DDOS.

I don't think so. YAML by itself doesn't have any entity expansion or include capabilities -- you'd have to rely on extensions or something else that isn't on by default. The reference (alias/anchor) mechanism just rebuild the serialized graph, so there's no expansion going on there. That said, I'm sure there are quite possible implementation issues, like most software.

To some extent every protocol which does not transfer message and payload size on a fixed offset in header can be called "crazy" as being vulnerable to all the problems with live parsing, terminators and unpredictable memory requirements.