It would work better if you did a local optimization after every crossover and ditched random mutation. I think they call it a memetic algorithm (http://en.wikipedia.org/wiki/Memetic_algorithm).

So take your every triangle vertex (x,y) and differentiate against the fitness (using finite difference), then gradient accent on that parameterization. That will let the triangles fit the edges effectively (I am thinking Mondrian here primarily)