I don't think it's viable to add this to the streaming engine, since it's a very limited use case.
If you use a sample start mod value that is not too high, you might still get the RR effect you need without wasting too much memory.
For example, if you set the SampleStartMod value to 16384, you have almost half a second of different material, which might be enough to kill the machine gun effect.
Also, if you use HLAC, memory consumption will drop by 50% (because it uses 16bit for the preload buffers).