R8 — NVMe sparse cache

Status: last. Do not implement until R5 watch is live, R3 pack is measured, and R7 has at least one cold/warm MEASURED row. Gate: working-set 10/50/100GB/1TB hit curve. RAM index + NVMe blocks + Swift backing.

This file is the contract, not the implementation. Sequence is in the 2026-08-28 plan: correctness → amplification → coherence → benchmark credibility → NVMe.