1. if using the fixed function pipeline, set D3DRS_NORMALIZENORMALS to TRUE

2. if using shaders, normalize() the normal after transformation, but before lighting.

3. a better alternative is to not scale at all unless absoultely necessary. If the scaling is there to make an object a specific fixed size, scale the vertices once when you load instead of though the world transformation.

4. mathematically, at the core of the lighting equation is L.N, a dot product between the vertexToLight vector (L) and the vertex normal (N). The product has the property that "I = L.N == cos(theta) * ||L|| * ||N||" which means "the light/illumination is the same as the cosine of the angle between the L vector and the N vector multiplied by the length of the L vector, multiplied by the length the N vector". If L and N are both normalised, their lengths are 1 so you just get "cos(theta)". Scaling the object scales the length of N when the vertices get transformed.