Agreed that option 1 is not scalable because the requester is often not interested in the embedded children. JDO and JPA are meant to perform lazy loading of data on demand but there is probably still overhead when no embedded data is needed.
Option 2 may look like strictly relational but it does not have to be. If your entities are not locked together into entity groups via ancestor keys, then the records are unrelated and free of constraints. You actually get to determine the tightness of consistency in your entity design and your application code.