Quote: "Why wouldn't the memblock be smaller though? If your removing verts from the index... why wouldn't the size be smaller?"
Because like I said to create 100 triangles you must create a memblock with 300 vertices, once done you can later make these 100 triangles use less verts via the vertex indices, however the original memblock you used to create the memblock mesh will not have differed.
Quote: "because less memory is being taken up by the memblock, but apparently I'm wrong about that."
It's entirely possible that it uses less GPU RAM, and only sends information about all verts that are referenced via the index data, however that in itself could be slower to calculate, and for most cases the reduction in vertex data would be very minimal compared to any texture data you store on a GPU, plus when you make a memblock from a mesh from an object that had index data, you get the original memblock and not a reduced one with only referenced verts.
In DBP there are currently 4 main(popular) ways to create terrains, the first is the matrix approach, where your terrain is textured by a single image, and you essentially slice this image up and shift the UV data for each tile to make it appear as if each tile uses many images. The advantage to this method is that each tile can be very different and highly detailed, the down side is that you cannot weld all overlapping verts so it's usually slower than other methods, and any transitions between tiles must be done yourself. You can make a variant of this using shaders that has fully welded verts, however this has potential to be slower on lower end cards, but I've yet to test this.
Another method is to have a fully welded mesh but store up to 7 layer opacity values per vertex, then texture the terrain with 8 images, multiplying the new layer with the opacity value for that layer then adding it. The advantage to this method is that you can have a sand texture at the shores as well as grass on the land and they will automatically blend smoothly, however if you have 8 textures you're sampling 8 textures per pixel and doing 7 multiplications and additions which could get slow. Plus each vertex requires 7 bytes to store alpha values in(which isn't much of an issue).
The third method is similar to the above one only you use images to store the alpha values of each texture layer, this is both good and bad as it allows you to specify the alpha precision of each layer, so you can have higher resolution alpha detail than per vertex, but unless you use a texture atlas you're limited to 4 layers(including the base). Also for only 3 additional layers(excluding the base) you're doing 6 texture lookups per pixel, so it's probably better to use the 2nd option.
And of course Advanced Terrains use a different method too, they have two layers, one is a highly tiled detail map, with a colour map on top which stretches over the whole terrain, this is probably the fastest method out of the others. However colour maps are highly limited as you're only really changing the colour of the detail map, you cannot have a sand texture at the beaches and a grass texture on the land, only a generic ground texture tinted yellow at the beaches and the same tinted green on the land.