The wikipedia page on rotation matrices shows 1's on the diagonal.

I believe scipy replaces the 1's on the diagonal with

w^2+x^2+y^2+x^2

That makes them the same for a unit quaternion.

For non-unit quaternions, scipy's matrix acts as both a rotation and a scaling.

For example:

if you take the quaternion = 2 +0i+0j+0k.

The rotation matrix will be the identity matrix (with only a w term there is no rotation),

Scipy's matrix will be 2*identity, because in also includes the scaling.