Most likely paths to error when estimating the mean of a reflected random walk