For what's it worth, I'd prefer a fail-late approach. I would find it somewhat surprising if validation of method X fails because method Y is invalid (which is generally unrelated except that it is defined on the same type). For metadata retrieval this would mean that getConstraintsForMethod() suceeds for correct methods, also if another method is invalid, while getConstrainedMethods() would fail as soon as one illegal method is present.

If going for fail early, would an illegal method also constructor validation to fail? And normal bean validation?

The other question is, whether this really should be specified or could be an implementation detail. In case of the latter we'd have to retrofit the TCK to handle implementations using a fail-early approach.