You should always measure first before blindly following dogmas.

Then you follow it by doubting it without measuring:

I doubt that the choice of serialization format is what dominates processor time in this case.

Also note I was not just speaking of today, but with future additions and changes. You can't measure the future, but you can plan ahead and build upon a good foundation, that the small format can be used for extending it in future. That's the idea. Also to test what you suggest they have to build the entire system years and waste lot of time and effort.