PS - I think it's perfectly reasonable to do something to improve Sqoop's precision w.r.t. nullability and typing. But I don't think it's part of this issue. I think that this issue ("the Avro output format is currently broken") is separate from "The Avro output format could be more precise."
It's also worth considering that since clients of these data files will use the Avro generic API (as Sqoop doesn't generate a schema file you can turn into a SpecificRecord), end-users will be dealing with object references returned by GenericRecord.get() anyway (as opposed to primitive types), so providing a more precise schema on the nullability doesn't buy you too much.