When would it make sense to use a non-default surface convection algorithm?

edit

This is only a partial answer, but as far a speed goes, any algorithm that doesn't involve a $ \Delta T$ term will allow E+ to run faster on average. The models that include a $ \Delta T $ term are accounting for buoyancy driven natural convection. For a given surface, the convection contribution to the energy balance equation is $ Q_c = h_c \Delta T $. This is a linear expression only when $ h_c $ doesn't depend on temperature, and linear equations can be solved with one iteration. Also equations that are near-linear can be solved faster than highly nonlinear ones. So a model with $ h_c \sim \sqrt { \Delta T } $ will generally run faster than one with $ h_c \sim \ { \Delta T } ^2 $. This also relates to the robustness because a more linear model is more likely to help the heat balance convergence.