Prescience (understanding your future requirements) can help dictate what your priorities are as a business, what features and flexibility to deliver today, but it has little to do with architectural soundness, which is simply a property of a system (independent of and not relative to the future).
This is my point -- to argue that a architectural sound choice is too expensive today is right only if the right skills aren't present. This is a problem for our industry to solve.
So the purpose of architecture, I'd say, is to first ensure soundness (which should cost no more than building an unsound system so long as you have the right skill set) and only then strike the right level of flexibility in the architecture (another, separate skill set that is valuable but less of a current industry problem, imo).
You can evolve the flexibility of a sound architecture but you can't do much with an unsound one. Inflexiblity in a system is generally a tractable "problem". Unsoundness is not (unwinding coupling, eg, is rarely tractable.)
But seeing into the future is risky/imperfect. So code bases are full of clever features and libriaries that are used exactly once. And where rapid growth happened but the code wasn't prepared, sketchy chains of conditions and runon code to paper over the gaps.