I will explain why that scenario should not occur and why we need detached entities.
Consider you are in a JTA transaction (JPA requires support for it) and fetch a.
Now you can call a.getB() either (1) in this transaction (i.e entity a is managed) or (2) when a is detached.
Scenario 1: now depending on your transaction isolation level, you might see or might not see what other transactions do. For example, if you have the SERIALIZABLE isolation level, then you will successfully fetch a.getB(), even if that row was deleted in a concurrent transaction. If that row was already deleted and your transaction sees that, it means that either your DB is inconsistent (no foreign key) or that you used the wrong transaction isolation level.
Scenario 2: the entity a is detached. When a LazyInitializationException is thrown, that means to me that you called a.getB() too late in order to guarantee a consistence in your application (as a is not managed anymore). In order to solve the problem you simply call it earlier when the entity is still managed. A NPE cannot occur.
Why we need the DETACHED STATE? Well, we need a state in which the changes to an entity instance are not tracked. Why?
Example 1: suppose you receive an entity (with persistent identity) in the EJB layer and that there were no detached state (meaning all entities should be managed). But we need to do a validation before persisting the entity. If that entity would be automatically managed, its changes would be automatically persisted to DB. So this new state was introduced.
Example 2: you receive in the EJB layer an entity, any you need to update ONLY 5 fields of 10 from that entity. If that entity would get automatically into the managed state, all 10 fields would be persisted. The solution in this case is to fetch a managed entity and to update the 5 fields ONLY in that entity.