I guess this relates to how we do in the temporal domain, where (I think) we normally interpolate the model result to fit the observation. I guess, because the observation is considered "a data point of truth" and also because the model normally has a higher temporal resolution than the observation. For a vertical profile I guess both observation and model result could have the higher resolution (although models become increasingly finer), but none-the-less I would say that your item 2 would be the most intuitive way to go.