The example at https://www.w3.org/TR/rdf11-mt/#dfn-lean
For example, the graph
ex:a ex:p _:x .
_:y ex:p _:x .is not lean, but
ex:a ex:p _:x .
_:x ex:p _:x .is lean. Ground graphs are lean.
In the question examples subject and object are flipped but otherwise G1 and G2 are the "RDF 1.1 Semantics" examples.
G1 is not lean.
You can remove "_:X foaf:knows _:Y" (or remove "_:y ex:p _:x") - that triple adds no new information. We know "_:X knows something" from the first triple "_:X foaf:knows ex:bob" and that is all the second triple says.
Graphs G3 and G4 are lean.
If we have G5: _:X foaf:knows _:Y then G5 is lean but G3 merge G5 is not lean.
G2 is lean because the triple "_:X foaf:knows _:X" adds the information that there is some resource that knows itself. The first triple does not say that.
Now I get it, Thank you very much.