Sequence primitives and per-stage neurons (WS-C)

Per-stage heterogeneous neurons

Every neuron stage can now choose its own kind, params, and surrogate. The presets accept additive neurons (per-stage kind) and stage_params (per-stage beta/threshold/reset/surrogate) maps, and TrainConfig gains additive stage_neurons/stage_params. A default build is byte-identical to before, and fc_legacy keeps its _fc1/_lif1/_fc2/_lif2 contract.

New stage kinds

topology/kinds.py grows conv1d, maxpool1d, maxpool2d, embedding, layer_norm, batch_norm, dropout, positional_encoding, attention, and multihead_attention, each with a module factory and an explicit NIR contract — mapped, passthrough (dropout is identity at inference), or unexportable.

sequence_mlp and sequence_attn

• sequence_mlp is built only from NIR-mappable kinds over a [T, B, L, D] sequence and validates end to end. Its neurons default to reset="zero", rendering a single nir.LIF with no Delay, so it is runnable by the Norse target too.
• sequence_attn is the spiking-transformer-shaped demo (embedding → positional_encoding → multihead_attention → layer_norm → linear → neuron). The installed nir has no embedding, attention, or normalisation primitive, so export raises the typed UnsupportedStageError naming the first unexportable stage. It stays available for simulation and introspection — the honest alpha precedent, applied to stages.

The toy token task (data/sequence_source.py) supplies [T, B, L, D] frames, and the client StageNeuronEditor edits the per-stage configuration. This enables sequence/attention experimentation, not production LLM training.