As mentioned I could not find any sources stating how to determine the subject in an ergative language. Since nobody else could provide them, I will go with what I already reverse-engineered out of existing grammars.
For example, Dyirbal is said to be OSV. When looking at an example it also is PAV. Hence S = A, making my example OVS/SOV. (Same argument can also be made e.g. for Basque SOV => APV.)
Now I will go forward with that, even though it is rather dissatisfactory, since it makes the concept of subject ill-defined in general and difficult to apply to ergative languages, since it is effectively defined as "whatever is considered the subject in English":
- syntax has no bearing on subjecthood: in nominative languages the agent of a simple transitive verb can omitted, whereas it is the patient in ergative languages.
- confusing behaviour with intransitives: in ergative languages we have the intransitive subject marked like the transitive patient, yet we still call the transitive agent the subject.
- semantics has no bearing on subjecthood: in English the patient can be promoted to subject by passive construction (whereas the antipassive construction of ergative languages has no effect on subjecthood).
- morphology has no bearing on subjecthood: in ergative languages the transitive subject is marked whereas in nominative languages it is unmarked.
I would much prefer to say that the subject is what is understood to co-refer if omitted:
- syntactically the subject would remain constant: in both ergative and nominative languages the subject could be omitted, but in the former it would be the patient and in the latter it would be the agent.
- consistent behaviour with intransitives: in ergative languages we have the intransitive subject marked like the transitive patient, and would indeed call the transitive patient the subject.
- semantics has no bearing on subjecthood, but antipassive in ergative languages would just like the passive in nominative languages delete the subject and promote the object - the difference being that the subject was patient in the former and agent in the latter.
- morphology would tend to correlate with subjecthood: both in ergative and nominative languages the subject of transitive verbs with default alignment would be the least marked case.
But as mentioned that does not seem the usual use, so I have to make do with that.