Note that the security of a whole system is harder to get right than the security of any one component; keeping it obscure just makes it much less likely that a white hat will notice the flaw and notify you. A simple example of this is the freshman CS majors' perennial idea that combining multiple PRNGs will yield a "more random" algorithm (it typically makes the result easier to predict). There are plenty of cases of a secure algorithm being used to build a system that ends up insecure because of some obscure flaw that nobody noticed initially.
I think that overstates the case. Your engineers don't have to be better, just good enough know what's correct and safe, or do research until they know.
My approach is to just add things that are clearly correct to published and studied systems that are thought to be correct. In this case, we're talking about applying multiple or multiple runs of one way hashing functions; I'd actually do some research before trying the latter, but it strikes me as safe, and the former is by definition safe, right? Whereas it would never even occur to me to combine PRNGs ... or not use a hardware source of entropy to begin with in the first place (they aren't that expensive in the scheme of things).
LinkedIn was not adding any sort of reasonable defense in depth security through obscurity, but foundational gross negligence.