I feel that outside of extraordinary circumstances, a test program should be added to show that a subsystem or driver at least *compiles*, if not prove that it works.
Option 1 is clearly a good approach and indeed in the context of the patch sets referred to above, this has happened. However it is not always possible or practical...
Under what circumstances is adding a small sample to at least build against the relevant APIs not possible/practical? Obviously it's not always feasible to prove that something works correctly, especially an external peripheral which may require some elaborate test apparatus, but proving that something at least compiles seems trivial to me.
The trigger for this thread is pwm/dw where no board currently supports the driver, it does not currently build but there is an expectation that the driver will be useful in the future.