First of, it is most often called MVP (model->view->projection), which is, tranform your model into world space, then into camera space and then into projection space (btw. a cube, not a flat 2d panel ;-) ).

But back to the issue you have. The basic idea of hom. matricies in game dev is, to move/rotate your object or single vertices in the world. The benefit of using matricies is, that you can chain actions by multiplying it with matrices, eg:
1. move object to enemy position
2. rotate object, so that it face enemy
=>
create translation matrix T
create rotation matrix R
=>
transform = T * R

So, the latter is the transformation matrix most often associated with a single object. It contains rotation and position (and scaling) information.

To transform a matrix, you just need to multiply the matrix with this transform:
output_matrix = transform * input_matrix

Well, if you want to transform a vertex, you only need the position, a vertex does not have a rotation or scaling, therefor you just need a vector (in other words a 1x4/4x1 matrix). So, you need to do this to transform a vector
v'=transform * v

But it is important to expand a 3d position to a 4 component vector, your vector should look like:
v = (x,y,z,1)

Sometimes it is only necessary to consider the rotation of a transformation, e.g. a normal, in this case your vector should look like this:
vn = (nx,ny,nz,0)
the 0 results in ignoring the translation part of the transformation.

World matrix (actual vertex pos is 5,5,5

The world matrix is not a vertex...

I would have though the new x & y co-ords

You haven't transformed any vertices yet. You can't have new x & y coords if you don't even have original/input x & y coords.

After creating a WVP matrix, you can multiply the vertex position [x,y,z,1] against the WVP matrix to get [x',y',z',w']. The screen (NDC) pos is then: XNDC = x'/w', YNDC = y'/w'

Hmmmm. Even when I post multiply the vertex vector (5,5,5,1) with the WVP matrix I get the exact same answer (as I suspected earlier).

1.66292 0       0       0
0       2.21723 0       0
5       5       16.002  1
0       0       -2.002  0

This doesn't make sense -- a vector4 (AKA 1x4 matrix, or 4x1 matrix, depending on conventions) multiplied by a matrix4x4 produces a vector4 result -- it does not produce a matrix4x4 result.

You want to be doing this, where P/V/W are Mat4x4's, and Vertex is a Mat4x1 (aka Vector4 - and Vertex.w should be 1.0):
ProjectedVertex = Projection * View * World * Vertex

Note that you can do it this way, which is three "Mat4x4 * Mat4x1" operations:
ProjectedVertex = Projection * (View * (World * Vertex))
Or you can do it this way, which is two "Mat4x4 * Mat4x4" operations, and one "Mat4x4 * Mat4x1" operation:
ProjectedVertex = ((Projection * View) * World) * Vertex
Both should produce the same result.

So... as well as your Mat4x4 Multiply(Mat4x4 a, Mat4x4 b) function, you also need a Vec4 Multiply(Mat4x4 a, Vec4 b) function.

Also, assuming there are no transforms, rotations, & scaling required. Couldn't you just multiply VP with the vertex vector?

Yes. The "world" matrix is actually the "model space to world space transform". If the vertices are already in world-space, then this matrix would be identity, which does nothing (so could be optimized out).