Strata: heavy models beyond VRAM

Strata is Mind’s layered model memory. One model appears fully available while weights live across three tiers:

Hot   → GPU VRAM
Warm  → System RAM
Cold  → NVMe (~/.buraaq/strata)

You type buraaq ai serve BIG-MODEL. Strata prints the residency map — never silent OOM-only failure when GPU+RAM+disk can place the model.

Plan only

buraaq ai strata Qwen/Qwen3-32B
buraaq ai strata Qwen/Qwen3-32B --strata gpu-ram-disk

buraaq ai serve MODEL --strata auto
# auto | off | gpu-ram | gpu-ram-disk

Doctor shows the cache path and free disk:

buraaq ai doctor

Override cache with BURAAQ_STRATA_CACHE.

Strata orchestrates peer engines. It is not a native Buraaq tensor pager yet.

Backend	Behavior
llama.cpp	`-ngl` + mmap; best path for disk tier
vLLM	CPU offload only for gpu-ram; no fake disk
Remote OpenAI	Strata off — remote owns memory

Cold tier slows decode. Speed class is printed: gpu-bound / ram-bound / disk-bound.

Next: Serve models with --strata auto, or back to Mind.