It is the nature of what a leak is, and why they exist. The compiler needs to determine what is the inside and what is the outside of a level, and it does so by looking at the entities.
Enclosed space that contains entities? World space, compile it.
Enclosed space that contains no entities? Void space, cull it from the map. (even if it doesn't touch the "outside", if there are no entities it is still void)
When it finds an entity in unenclosed space, it gets confused and the resulting leaf structure cannot be used for visibility calculation.
Brush entities exist as two separate but associated components. One is the entity, which exists at a specified location in the map. The other is the brush, which is stored away from normal brushes as a "model". The entity half specifies (by number) which model it is associated with (you can see this if you look at the entity data in a compiled BSP). If you -onlyents to remove a brush entity, you will see the disastrous results from this, as the model data is not updated but all the entities are, thus shifting each brush entity to be associated with a different model (trigger brushes as solid doors, doors being death pit hurts).
Where I am going with this is: only the origin needs to be inside because that is actually where the entity exists, it does not "fill" the infinite number of locations within the brush. The brush and the entity are not one in the same, that's why it doesn't leak.
