ECPs might be nice, but the worst deficiency IMHO is the lack of robust convergence accelerators. There's DIIS and SOSCF, but the latter doesn't seem to work with DFT.
I'd recomment implementing ADIIS [JCP 132, 054109 (2010)], it's very easy to implement and is able to converge a lot of systems where DIIS fails.
Another worthwhile feature would be a gradient descent minimizer for cases where Roothaan-Hall does too jumps that are too big, or where degeneracies are an issue. I guess GDM [Mol. Phys. 100, 1713 (2002)] is still the best(?) method.