ADR: StoryLM-0 host train loop

Context

Section 23 can roll 32 tokens from a patterned fixture. Those weights are not fit to any corpus, so greedy text is an artifact of the recipe. Section 15 already performs one sum-NLL SGD step on every trainable decoder weight.

A train primitive would invent an optimizer ABI and a data loader. Mean NLL is still unspecified. Decoder-wide AdamW is Section 25.

Decision

Specify a host train loop as the existing composition in 24-storylm0-host-train.md: 40 epochs of Section 15 SGD over three sliding windows of once upon a time the end, init from the patterned fixture, then Section 23 greedy. No new operator. Do not add mean NLL or AdamW in this slice.

Consequences

• just validate-train can show corpus NLL drop and a locked greedy cycle without a new compiler operator.
• Visitors can see a taught continuation; they still cannot train a useful language model or tokenize real text.
• Mean NLL is Section 26. The trained 32-token fixture is Section 27. Decoder-wide AdamW is Section 25. Remaining opens include top-k / top-p, in-place write, and a real corpus.