Assuming that (y, x1, mediator, x2) is the true model you have omitted variable problem when estimating the model (y, x1, mediator) which very plausible creates bias see wiki on: Omitted-variable_bias and makes the OLS estimator inconsistent. This bias and inconsistency is one argument for not starting with a minimal model and adding variables but instead starting from af maximal model - including all variables, interaction etc. - and then reducing.... since the OLS estimator will still be consistent if you estimate a model including variables not included in the true model.