RotorQuant outpaces TurboQuant with Clifford rotors

This feels less like a flashy benchmark stunt and more like a systems paper that can actually shave real inference cost. The caveat is that it wins by changing both the geometry and the kernel shape, so the real test is how broadly those gains survive outside this KV-cache setup.

• –The 44x parameter drop is the real deployment story, not just the speedup, because KV-cache compression is usually memory-bound before it is FLOP-bound.
• –The fused-kernel claim is believable: tiny 3D rotor blocks keep work in registers and cut the memory traffic that makes dense matmuls expensive.
• –The QJL-corrected validation is the key proof point, since matching real-model attention fidelity matters more than synthetic-vector MSE.
• –The tradeoff is that RotorQuant changes the statistical assumptions TurboQuant relies on, so it may be narrower than the original method outside KV-cache compression.
• –If the implementation holds up, this is a strong example of geometric algebra turning into practical inference engineering.