A CANONICAL ANSWER · FROM THE VINDEX KNOWLEDGE GRAPH
Operation family does not imply state shape
An operator may declare more than one continuation region, and being 'attention' or 'a recurrence' decides nothing about how many. Both recurrence families carry a convolution history beside their folded state — and the hybrid witness's conv-QKV attention, recognisably attention, declares a per-position KV cache AND a convolution history on the same layer. The regions are declared by the operator and consumed by execution; a provider that can hold only KV rows refuses such a layer rather than silently allocating half of it. That refusal is the fail-closed philosophy reaching all the way into runtime API design.
SUPPORTED · witnessed — mamba2attn-250m: KV + conv history on one layer, batch-vs-decode bitwise across the step boundary