We can live with that.
Normal projection for decal is dependent on decal rotation. Suggest skipping tangent space normals, and outputting world space normals by rotating result of normal map sample based on difference between world up and decal up vectors.
Well, yes.
Indeed, you can use partial derivative of underlying position to estimate normal, that you would use to blend between projections.
Generally, I don’t agree here, that it is game breaker. It most cases it is safe to leave it as it is.
But this issue is tied together with another one. You really need those underlying normals anyway to get rid of lit backsides of the mesh, where the decal rotated normals to point towards the light, while it should be fully shaded.
I will revert with clarification a bit later.