Replace fresh full-attention storage before prefill with pages storing what
the composition's scheme kv names for each layer (bf16, affine at the
layer's width, TurboQuant read with the layer's composed read), sized for
capacityTokens (the pool grows past it). Pages store their format from the
first position. Sliding layers retain their layout; pre-warmed caches retain
their existing storage. Nothing changes without paging.
Replace fresh full-attention storage before prefill with pages storing what the composition's scheme
kvnames for each layer (bf16, affine at the layer's width, TurboQuant read with the layer's composed read), sized forcapacityTokens(the pool grows past it). Pages store their format from the first position. Sliding layers retain their layout; pre-warmed caches retain their existing storage. Nothing changes without paging.