Gemma-4-class models could not QLoRA-train with offload=True at all -- the first
backward raised "backward re-dequantization read an offload-evicted expert". The
evict post-hook fired during the gradient-checkpoint recompute and un-staged the
layer before its own backward read it; offload.py documented the opposite as an
invariant. OLMoE, Qwen3-MoE and GraniteMoe stop the recompute early, which is
why the wrong premise survived.

Two new ways to fit a model that does not fit: nvme_experts serves the cold
expert tail from NVMe, and dense_offload/dense_disk serve the 114.4 GB dense
side straight from the checkpoint's safetensors, byte for byte.

The flagship matrix's B1 bit-exactness gate is WITHDRAWN as vacuous -- it hashed
offload placeholders and compared sha256(b"") with itself -- and its B2 table is
corrected to the metric the protocol registers. The performance numbers are
independent measurements and stand.

CHANGELOG backfilled for 0.6.5, 0.6.6 and 0.6.7, which shipped with no record.