Better document that initial guesses must be strictly feasible with respect to the inequality

[josyoun]

We need to better document that the initial guess to the optimization must be strictly feasible with respect to the inequality. I thought this was already in the manual, but I can't find it, so we need to update the manual in either case.