Confusion about test accuracy

On https://www.mitchelloharawild.com/blog/fable/

Mr. Ohara-Wild writes in his blog: '[...] The in-sample accuracy suggests that the ETS model performs best. This is because it has the lowest values for all accuracy measures (lower values indicate less errors). [...]'

If I compute row means from the resulting object of accuracy() and arrange it descending, I get an opposite ranking. Models manually top ranked- are then poor for fable (in- and out of sample). What is the reason for that?

Are there any additional, internal criteria for the ranking?

bestModel<-fabletools::accuracy(myModels)
            
msg<-paste('Best Model Fable:')
print(msg)

info<-bestModel %>% 
	dplyr::rowwise() %>% mutate(myMean = mean(c(ME,RMSE,MAE,MPE,MAPE,MASE,RMSSE,ACF1),na.rm=T)) %>% 
	ungroup()
print(info)

bestModel<-bestModel %>% dplyr::arrange(myMean)

msg<-paste('Best Model Ich:')
print(msg)

print(bestModel)

Start with the accuracy() output and the fabletools model-selection behavior described in the issue, then compare both rankings against the linked blog explanation. Done means identifying whether the discrepancy comes from the ranking criteria or the example and documenting the reason clearly.