tiny_prefill_decode_npir

Tiny autoregressive prefill/decode NPIR fixture for multi-graph GQ tests (NPP02-7270). KV cache is graph I/O (not an internal stateful buffer), matching the USER_INPUT/USER_OUTPUT/USER_INPUT_MUTATION form measured on a real smollm2 torch.export graph_signature. Deterministic (seed 0), 2 layers, reduced dims.

• prefill/: input_ids [1, 4] -> logits [1, 4, 16] + 2 K/V pairs [1, 1, 8, 8]
• decode/: input_ids [1, 1], position_ids [1, 1], 2 K/V pairs (same shape) -> logits [1, 1, 16] + the same K/V pairs, updated in place

Composition edges (one per K/V tensor, [k0, v0, k1, v1, ...] order):

• prefill.index_copy_.output_0 -> decode.k0
• prefill.index_copy__1.output_0 -> decode.v0
• prefill.index_copy__2.output_0 -> decode.k1
• prefill.index_copy__3.output_0 -> decode.v1

No decode -> decode edge is declared: the autoregressive feedback (decode's own K/V output feeding its next-step input) is a cycle and is correctly rejected by multi_graph.composition.topological_order (CyclicCompositionError) -- it is runtime/wiring's concern, not GQ's calibration DAG.

Calibration payload gotcha: decode's two non-edge inputs (input_ids, position_ids) need a PT-format payload (torch.saved dict[str, list[torch.Tensor]]) whose per-sample tensors already carry their own batch dimension (e.g. shape [1, 1], not [1]) -- unlike NPY, the PT loader does not add a batch dimension.