TurboQuant ARM port stalls on Android

This is the classic gap between a strong research result and a shippable runtime feature: the math is real, but the integration work is still missing. The experiment is valuable because it separates "can compile on ARM" from "can actually run TurboQuant on a phone."

• –Google’s release backs the headline claims: 3-bit KV caches, at least 6x memory reduction, and up to 8x speedup on H100s.
• –The Android result suggests the current llama.cpp path is still missing the quantization type registration, so a successful binary build is not the same as feature support.
• –That matters on 8GB phones, where a real KV-cache compression win could be the difference between workable long context and out-of-memory crashes.
• –The build failures also highlight the usual mobile-port landmines: NDK toolchains, stray x86 flags, and target plumbing that desktop-centric ML code often assumes away.