1) Normalisation at all costs is foolish - if the cost exceeds the value, then don't do it. That isn't complicated. Denormalised data sometimes points at design flaws, but even then all systems have design flaws and they don't automatically need to be fixed. Quality is expensive, like every other property (even doing things the cheap way is expensive, ironically - software is all about managing costs).
2) For any given user it is better to have denormalised data where the data model is perfectly aligned to their use case. For a system with multiple users it is better to have normalised data. And the corollary is that any data important enough to be recorded is probably valuable enough that it will eventually have multiple interested users even if the person building the system swears that this time is different - so they should normalise their data. Brownie point to anyone who has reached enlightenment and understands the you of 12 months hence is a different user with different needs of the data.