> hard to imagine a case where a 623-dimansionally equidistributed PRNG could fail
Well, there is the generalization and improvement of MT19937 in form of the WELL generators. They go beyond that, so there probably are cases where it's needed.
> PRNGs could become obsolete but "truly random" will always mean the same
I'd argue that far fewer people need truly random than those who just need a dice roll.
There are those who work on crypto. They won't need anything in the C++ stdlib because there are libraries specifically for that. They also provide much more that is needed for programs utilizing cryptographic primitives. This group should know what they're doing and choosing, and if not should have no business writing such software. And they won't even accidentally pick MT.
There are those who do numerical simulations. They neither need or can use a "truly-random" generator because submitting a paper with the words "to reproduce our results, grab the following 2 TiB file of random bits ..." is probably frowned-upon. Reproducability of a sequence is a feature, and a good one. Heck, quasi-random algorithms also have their place. In any case, you shouldn't do simulations with just a single PRNG either to rule out interactions between intricacies of your model/simulator and the PRNG. So people in this group are likely to use a framework or library that caters to their needs, too, which includes several different PRNGs (even obsolete ones for checking older results), distributions, special data structures needed for certain kinds of simulations (e.g. event queues), etc. This group also should know what they're doing and why, otherwise they shouldn't publish research at all.
And then there is everyone else who just needs a random number every now and then. Maybe for shuffling a playlist, maybe for rolling a die. Having easy access to a PRNG that works well for a large number of use cases (they wouldn't have any idea what to choose anyway) is a major benefit to this group. They don't care about (or notice) the difference between pseudo-random and truly random (even though the latter sounds more impressive and the former somehow not random enough, even though anything that doesn't look random in the former case is often bad seeding, e.g. in a loop). They just want something to work. MT19937 is a very safe choice for this group, and as an added benefit, it's much better and often faster than LCGs.
Well, there is the generalization and improvement of MT19937 in form of the WELL generators. They go beyond that, so there probably are cases where it's needed.
> PRNGs could become obsolete but "truly random" will always mean the same
I'd argue that far fewer people need truly random than those who just need a dice roll.
There are those who work on crypto. They won't need anything in the C++ stdlib because there are libraries specifically for that. They also provide much more that is needed for programs utilizing cryptographic primitives. This group should know what they're doing and choosing, and if not should have no business writing such software. And they won't even accidentally pick MT.
There are those who do numerical simulations. They neither need or can use a "truly-random" generator because submitting a paper with the words "to reproduce our results, grab the following 2 TiB file of random bits ..." is probably frowned-upon. Reproducability of a sequence is a feature, and a good one. Heck, quasi-random algorithms also have their place. In any case, you shouldn't do simulations with just a single PRNG either to rule out interactions between intricacies of your model/simulator and the PRNG. So people in this group are likely to use a framework or library that caters to their needs, too, which includes several different PRNGs (even obsolete ones for checking older results), distributions, special data structures needed for certain kinds of simulations (e.g. event queues), etc. This group also should know what they're doing and why, otherwise they shouldn't publish research at all.
And then there is everyone else who just needs a random number every now and then. Maybe for shuffling a playlist, maybe for rolling a die. Having easy access to a PRNG that works well for a large number of use cases (they wouldn't have any idea what to choose anyway) is a major benefit to this group. They don't care about (or notice) the difference between pseudo-random and truly random (even though the latter sounds more impressive and the former somehow not random enough, even though anything that doesn't look random in the former case is often bad seeding, e.g. in a loop). They just want something to work. MT19937 is a very safe choice for this group, and as an added benefit, it's much better and often faster than LCGs.