Continuation state generalises KV

Because KV is one state family, not the definition of model continuation. The model program declares what persists between tokens: softmax attention declares KV rows, MLA a latent-compressed cache, KDA and Gated DeltaNet a fixed-size recurrent state, Mamba2 two regions — an SSM state plus a convolution history. A pure-SSM container describes its whole continuation with no KV row anywhere — and the runtime reads that declaration from the plan, never from an architecture guess.

PASSED · mamba2-780m: continuation reported as recurrent state only — 18.9M elements, constant in sequence length