Should I denormalise narrow, repeititve dimension tables

[Snowlock]

I have read that narrow dimension tables containing repeated data (e.g. order status, that kind of thing) should be denormalised onto the relevant fact table - that Vertipaq will optimise this using dictionaries. However, the article is about five years old, and I'd like to ask what the current best practice is. Particularly if you then filter on or write a measure against a denormalised column, will that then do a string comparison on each row in that column? Or will it find a key from the dictionary, and use that?

I guess I'm asking - is it better to add a denormalised column to the fact table, or a completely new, normalised dimension table?

EDIT - hmmm it's from Jan 11th, I guess it's not as old as I thought:
https://towardsdatascience.com/how-to-reduce-your-power-bi-model-size-by-90-76d7c4377f2d

[amitchandak · Accepted solution]

@Snowlock , It is always better to be in Star schema. If a dimension has levels in different tables merge those.  to make sure only 1 level join with fact.

https://www.sqlbi.com/articles/the-importance-of-star-schemas-in-power-bi/

A dimension can be role-playing or sometime you need more than one copy of dimension as per need

If a dimension does not add any value separately, means it belongs to only one fact and use only as a filter or slicer it can be merged with the fact