I think both methods can work well. But Fogman is right when it comes to the point of learning how far you can go until you reach limits. Because of that he already wants to use textues and shaders in a prototype to see how it works, how it scales, how detailed he can go later on.

I agree with that. You dont need the prettiest textures for that, but you need textures, normal maps, spec maps, maybe even displacement and opacity. You need all this for testing.
And when you texture everything with some kind of a measurement pattern displaying 1 meter or similar, then you will not see how many draw calls a final level will create. You will not experience how many draw calls will bring your game down, how you have to optimize, to copy several textures into one, to group models, whatever.

Because of this he suggested to work as close as possible to a final result. And it indeed makes a lot of sense.