First, let's talk about what you're concerned about that is not an error:
The reason your normal map is all seamed up is because it has seams.
A tangent-space normal map is measured relative to the surface normal and the UV tangent, which is the direction of increasing U. The direction of increasing U changes across seams. Because of that, don't expect the color of a normal map to be continuous across seams. It's not supposed to be, it doesn't have to be. Instead, look at the actual normals:
We can see, with a roughness 0 glossy shader, that the normals are continuous across that region, even though the color of the normal map isn't.
You may wonder what I'm doing with my normal map there. Your pictures indicated you baked this for DirectX, while Blender is an OpenGL application. The significance of that difference for baking normals is the handedness of the coordinate system, and to convert a normal map from DX to OGL or vice versa, you invert the green channel.
That doesn't mean there are no concerns here. Let's look at some other parts of the normal map:
This kind of thing doesn't look good. Although you didn't provide the high poly, it's easy enough to imagine what happened here. Rather than the low poly's faces following the high poly's, giving you smooth, continuous normals, you've made a low poly out of self-intersecting geometry. You're not going to get the smooth normals of your high poly on any of these places made out of self-intersections.
Now let's look at your cage, which you provided in the file:
Your cage appears to just be your low poly, with a live displace modifier. There's not really any reason to use a cage in that situation (as you've noticed, using it doesn't change anything), because a bake with some ray extrusion, or whatever SP calls it, is the same as baking from a displaced mesh. However here, we can see some self intersection in the low poly, which is going to translate into some wonky reads from the high poly. If you wanted to use it, it would be smart to edit this cage to smooth that area out with alt-s on individual vertices.
One last thing that worries me here, although there isn't enough information to absolutely know that it's a problem: this mesh is all quads. Isn't that supposed to be a good thing? It's not, when you're dealing with multiple engines (Blender, SP, UE) that aren't necessarily going to triangulate meshes in the same way. Different triangulations will lead different interpolation of vertex data across the mesh, and whenever you bake any kind of texture, you need to know that you're baking it to a particular triangulation.
That's what happens when we flip triangulation on your model (fixed to fixed alternate.) Neither of these triangulation methods is necessarily what SP used for the bake. If you are using SP to paint textures, and UE to render, you should be exporting a consistently triangulated model to each.
Substance Painter materials view, how the textures (and lighting) are supposed to look... but if you zoom in you'll see the normal seaming issue.
