Interface PagedKvComposition

Positions per page.

The pages' read of at most eight queries without an array mask (decode, and prefill and verification windows that short): the direct Metal kernel over the blocks (Lab numerics) where true, otherwise the gathered blocks through the fused SDPA (the numerical control). Pages the kernel does not read (TurboQuant, affine widths that straddle a 32-bit word) read gathered. Pages the kernel reads have no compiled decode form (the kernel's partitions follow the context length): their composition resolves compiledDecode off.