Why aren't only those model versions that best represent the historical climate used in the ensemble?

Comparing model simulations of climate and climate change with those from observations is a recognised validation technique. All models have errors and even those with larger errors can still give useful information; there is no cutoff between good and bad models. For this reason, it is important to include all model versions in the ensemble.

In detail

Although validating against historical climate is a recognised way of weighting models, we have to make an assumption that the model weighting based on historical observations can also be applied to projections for the future.