I think the full answer is a little of both. When choosing to use CFD for any analysis there is always a trade-off between accuracy vs. cost.
Let's take the rocket nozzle for example. The real physics of the problem include effects such as unsteadiness, turbulence which is inherently 3D, combustion, reacting chemicals for the combustion products etc. These can indeed be simulated with today's computers but as you add physics and complexity the computational cost rises. This sets the limit on the cost side.
However, let's say we really just want/need to understand if the nozzle is over/under expanded. Well then a simple 2D axisymmetric analysis (which is far cheaper computationally) will suffice. If we do care about things like how the reacting combustion products affect nozzle efficiency then those physics need to be included and consequently the computation cost rises as our level of accuracy increases. This sets the limit on the accuracy side. That is, what level of answer is "good enough"?
Rounding back to your question, it is a bit of both where the level of analysis is the result of the balance of cost and accuracy desired.