One option in this case is to take the Wheeler-Feynman view that there should be no initial unsourced fields -- in other words, that all fields should be sourced by charged matter -- but that requires expanding your analysis to include the field sources, and it also requires more than just the free-field equations. More commonly, people just assume the Sommerfeld Radiation Condition that there are no incoming free fields at all, other than those with known sources.
More generally, note that most physicists treat the complete initial state of a modelled system as an externally controllable input to the "model" in question, so we are setting those values anyway when we solve a problem. Since we set them by hand, if they are supposed to obey some constraint, we just have to make sure that we set them according to that constraint.
If we don't treat the initial values as a external input, your question becomes a lot more interesting! For example, if you're solving an EM problem using action extremization, then some of the input to the problem is a future constraint, and a portion of the initial state is an output, solved by extremizing the action. In that case, for classical EM, I'd highly recommend changing the equations around to be in terms of the four-vector potential $A_\mu$ (or, equivalently, the scalar potential $V$ and the 3-vector potential $\vec{A}$), where the constraints you're worried about (on $\vec{E}$ and $\vec{B}$) become automatic. You still have a choice of gauge in this view, but you can use the freedom of choosing a gauge to not have to worry about constraining the initial data in any particular way.