We currently store feature-linked annotations in a different feature dataset than the features that the annotations are linked to. This is because we use a different projection for the annotations in our maps than the features themselves. I see in the documentation about annotations:
"A feature-linked annotation feature class inside a feature dataset should link to a feature class within the same dataset."
I'm just curious why this is required and what behavior should I expect if we don't store the annotations in the same feature dataset.
I suggest doing some testing on this first. The relationship set up automatically for feature-linked annotation is Composite. So you are asking that deleting a feature in one dataset deletes the linked annotation in the other dataset. This might work, but I would expect locks and editing rights on two FDSes to make this process messy.
When you reshape features, resize annotation fonts, or make many types of changes, the feature-linked annotation geometry changes. Annotation is complex geometry (polygons and points, I believe). So this may not matter in most cases, but I would think sometimes text could get cut off or become distorted when created in a different projection.
This is an interesting dilemma, so let us know what happens.