This is a very nice question.
First, I have to say there are some advanced power tricks (too complicated to get involved with here) that let you have your cake and eat it too: always make sure the premium model is more expensive than the base model, but accomplish that by not compromising the orthogonality of your experiment. If you worked with an advanced consultant, you could accomplish that with some power tricks within our software. I have nothing written on that I can share, if that is your follow-up question. Sorry. This topic has been discussed at "Turbo CBC" seminars in the past by David Lyon (the next Turbo CBC is going to happen in Amsterdam in October, BTW).
Next, the debate regarding having conjoint studies that occasionally show dominated concepts has been in play since the 1970s. In the 1970s, they worried about orthogonal arrays for card-sort conjoint producing some cards that were inferior (logically) to other cards within the same experiment. Researchers experimented with such orthogonal arrays versus other arrays (sometimes with prohibitions) that avoided dominated concepts (called Pareto efficient designs). Green and Srinivasan reported in their 1978 article that multiple research projects investigating this were inconclusive: both approaches worked about equally well. In other words, there didn't seem much harm in having occasional dominated concepts.
Pinnell and Huber looked into this for CBC studies in the 1990s and found very modest gains to be had by avoiding dominated concepts. Paul Johnson implemented another experiment along these lines and presented it to the Sawtooth Software conference in 2013. Same essential findings.