Each of these are just layers of control at different lifecycles of agent code generation. Analyzers are nice because it gives targeted, static analysis that can prevent certain classes of errors very early and at lower iterative cost (e.g. a build)
I would certainly consider Roslyn analyzers capable of covering "a tiny fraction" of possible errors :) They are quite capable of covering for many common types of structural coding mistakes.
Each of these are just layers of control at different lifecycles of agent code generation. Analyzers are nice because it gives targeted, static analysis that can prevent certain classes of errors very early and at lower iterative cost (e.g. a build)