ADR: StoryLM-0 deterministic `argmax`

Context

Phase 0 §8 lists argmax as a reduce op whose result is I64 regardless of input. The bootstrap compiler still rejected the name in reduce_ok. AGENTS.md step 4 asks for deterministic argmax after functional AdamW, then seeded sampling, then KV-cache.

The typed-rule box in Phase 0 writes Tensor[dt, …] for every reduce. The prose overrides that for argmax / argmin. Tie-break was unspecified.

A new argmax keyword would invent surface that already exists as x.reduce(op=argmax, axis=…).

Decision

Specify host argmax in 17-argmax.md:

• Result dtype is I64 (Phase 0 prose).
• Tie-break is lowest index (NumPy / PyTorch; strict ogt / sgt so equals keep the old index).
• Admissibility matches reduce max (floating + signed int).
• Host-only; device rejected at type-check.
• No VJP. argmin stays unimplemented.

Do not add sampling or KV-cache in this slice.

Consequences

• Greedy token ids type-check and lower on the host.
• Visitors can compile storylm0_greedy.kina; they still cannot generate text or sample.
• Sampling, KV-cache, device argmax, and argmin remain follow-on work.