ADR: StoryLM-0 prefill → decode handoff

Context

Section 20 starts the host loop from a self-attending seed. Section 19 left “slice or split the prefill Kᵀ / V into a cache” open. storylm0_forward.kina already computes those tensors and discards them. T_max scratch would invent an in-place write that Phase 0 does not define.

Decision

Specify prefill handoff as the existing composition in 21-prefill-handoff.md: the forward graph returns (Kᵀ, V, logits); the host slices the last logit row, applies argmax, and calls a static-S decode step whose past length equals the prefill T. No new operator. Do not add T_max scratch.

Consequences

• storylm0_prefill.kina and a .so numeric check land without a new compiler operator. Host extract_slice .so lowering now includes --lower-affine.
• Visitors can see a prompt cache handed to one decode step; they still cannot grow S inside while.
• T_max scratch is Section 22. The 32-token checkpoint test is Section 23. A prefill-T family remains follow-on work.