> A huge number of bugs are obvious to the eye, but have high odds of escaping tests. Doubly so in code where there's no "right" output to test for: as in the case of any non-optimal optimization function.

I agree in principle, but your example is off. Optimization is testable, and even provable. If your optimizers doesn't rely on ad-hoc techniques, you can also usually give an estimate of how far away your solution is from the true optimum.