Stuff that depends on the existing bugs in order to function can keep functioning. Everything else can move to something sane.
Obviously I'm missing something here.
Also, just, it seems unlikely that the kernel project is going to agree to run two entire unrelated CSPRNG subsystems at the same time! The current LRNG is kind of an incoherent mess root and branch; it's not just a matter of slapping a better character device and system call on top of it.
How is there any overlap in the devices that can't have something clever figured out and devices that could possibly see an update to their kernel code?
Note: Extremely good reason for breaking userspace is along the lines of "/dev/random has been found to be insecure causing mass security mayhem" not "man I'd really like to ignore the 0.01% of users this would cause an issue for so I can get my patch in faster".
For a start, there's a long tail of migrating all useful software to /dev/sane_random. Moreover, there's a risk new software accidentally uses the old broken /dev/random.
Besides, /dev/sane_random essentially exists; it's just a sysctl called getrandom().
but that risk is at 100% NOW, how is it not worth reducing it?
While there are platforms with better and worse hardware sources of unpredictable bits, the problem with Linux /dev/random isn't so much a hardware issue, but rather a software one. The fundamental problem Linux tries to solve isn't that hard, as you can see from the fact that so many other popular platforms running on similar hardware have neatly solved it.
The problem with the LRNG is that it's been architecturally incoherent for a very long time (entropy estimation, urandom vs random, lack of clarity about seeding and initialization status, behavior doing bootup). As a result, an ecosystem of software has grown roots around the design (and bugs) of the current LRNG. Major changes to the behavior of the LRNG breaks "bug compatibility", and, because the LRNG is one of the core cryptographic facilities in the kernel, this is an instance where you really really don't want to break userland.
The basic fact of kernel random number generation is this: once you've properly seeded an RNG, your acute "entropy gathering" problem is over. Continuous access to high volumes of high-entropy bits are nice to have, but the kernel gains its ability to satisfy gigabytes of requests for random numbers from the same source that modern cryptography gains its ability to satisfy gigabytes of requests for ciphertext with a 128 bit key.
People looking to platform hardware (or who fixate on the intricacies of threading the LRNG isn't guest VMs) are mostly looking in the wrong place for problems to solve. The big issue today is that the LRNG is still pretty incoherent, but nobody really knows what would break if it was designed more carefully.