I know that one shouldn't use the built-in PRNG for any 'serious' work...
And that's pretty much why the MS C built-in PRNG has never been improved. If they improved it sufficiently that it was no longer glaringly obvious that it is useless, then someone, and probably many someones, would try to use it for critical purposes and then blame them when it all came unravelled.
And whilst I've never seen it written anywhere, that might also be a good reason for Perl not providing a better PRNG built-in. If it is sufficiently better to be worth while including, someone is going to assume it is sufficient for purposes for which it isn't.
That said, it would be nice to have a good--in the simulation sense rather than the crypto sense--PRNG as a part of the core. Preferably such that if enabled through a use (or pragma), and seeded with a known value, it would produce consistant sequences across platforms. It would be a considerable boon for testing purposes.
The MT would probably do that, but would forcing the users to explicitly select it for use be sufficient to prevent the blame game when people use it for the wrong purposes?