nir unmappable decisions

NIR-unmappable stage decisions

Nine stage kinds have no NIR rendering in the installed nir package, so export raises the typed UnsupportedStageError naming the kind (stages_unmappable.py, neuron_nodes.py). This record gives each kind a verdict — await-upstream-primitive, decompose-into-existing-NIR, or permanently-simulation-only — with the reason and, for decompositions, the exact construction. The error message points here so a caller hitting it learns the status.

This is broader than the C runtime's own kind gap (Linear/Conv1d/ LIF/IF/CubaLIF, which export fine but the C loader rejects): these stages fail before a graph even exists. The C side is profiled in plans/nir-c-runtime-profile.md.

Stage kind	Verdict	Reason	Follow-up
embedding	decompose-into-existing-NIR	A row gather is exactly a one-hot product: expand token ids to one-hot at ingest, then Affine(weight=table, bias=zeros).	#169
maxpool1d	await-upstream-primitive	No max or comparison primitive exists. SumPool+Threshold is exact only for binary inputs, which the stage does not guarantee.	—
maxpool2d	await-upstream-primitive	No max or comparison primitive exists. SumPool+Threshold is exact only for binary inputs, which the stage does not guarantee.	—
layer_norm	await-upstream-primitive	Per-sample mean and variance are runtime reductions with no primitive, and input-dependent statistics cannot fold into static weights.	—
batch_norm	decompose-into-existing-NIR	Eval-mode statistics are frozen: fold per-channel gamma/sqrt(var+eps) and beta-gamma*mean/sqrt(var+eps) into one Affine. Exact for [B, C] inference inputs only.	#170
positional_encoding	await-upstream-primitive	The fixed table varies along the sequence dim and slices to the input length, so it fits no static weight; there is no Constant node.	—
attention	await-upstream-primitive	The Q/K/V projections already map to Linear/Affine; the score matmul, softmax, and weighted sum need dynamic-matmul, exp, and reduction primitives with no exact decomposition.	—
multihead_attention	await-upstream-primitive	Per-head Q/K/V projections already map to Linear/Affine; the score matmul, softmax, weighted sum, and head mixing have no exact decomposition.	—
alpha	await-upstream-primitive	Three coupled states (syn_exc, syn_inh, mem) with spike-gated reset across all of them: zero-reset needs a dynamic multiplier, subtract-reset terms contain dynamic products, LI has no reset, and LIF hides its analog state.	—

Until the follow-ups land, every kind above stays simulation-only: available for simulation and introspection, rejected at export.