Epic: C-side training support (zero-Python/zero-torch enablement)

Current state

nir-c-runtime is inference-only by design (README.md:77-79: "Training. This is inference-only."). Every training capability lives in spikeforge's Python/torch stack, and every bridge between the two projects flows one way: Python → C.

What C-side training would need (survey)

Mirrored from spikeforge's actual training path (spikeforge/training/):

• Weight updates: spikeforge trains by surrogate-gradient BPTT — ScaleUpMixin._train_batch runs cross_entropy + AmpController.backward (spikeforge/training/scaleup_mixin.py:86-101) through snnTorch neurons whose spike_grad surrogate is resolved from dynamically-discovered snnTorch factories (spikeforge/neurons/spike_grad.py:20-25, spikeforge/introspection/surrogate.py:71-79). The C runtime has no backward pass, no gradient tape, no loss, and no optimizer; its Threshold is a strict Heaviside with no surrogate hook (include/nir_c_runtime/ops.h:81-88, src/ops.c:435). The only torch-side local rule portable without autograd is pair-based STDP (spikeforge/memory/stdp_synapse.py:64-92).
• Dataset loading without torch: build_loader is torchvision + snntorch.utils.data_subset + torch DataLoader (spikeforge/data/data_loader.py:7-19); the dataset registry is torchvision classes (spikeforge/data/datasets.py:11-12,22-46). C has no loader at all, and tensors are batch-1 float-only (include/nir_c_runtime/tensor.h:1-6).
• RNG + spike encoding: rate/latency/delta/random codings in SpikeEncoder (spikeforge/encoding/spike_encoder.py:9-12,96-101) use torch.rand/Bernoulli sampling; seeding is torch.manual_seed/torch.Generator (spikeforge/tracking/seed.py:33-47, tools/export_fixture.py:49-52). C has no RNG and no encoder.
• Checkpoint save/load: checkpoints are torch.save pickles (.pt: state_dict + meta + history + manifest, spikeforge/network/model_store.py:34-54) with a manifest card (spikeforge/training/checkpoint_mixin.py:62-104). C cannot read or write this format, and its JSON module is parse-only — include/nir_c_runtime/json.h:43-74 exposes no serializer.
• NIR export of trained state: to_nir(spec, module) copies learned tensors out of a built torch module (spikeforge/nir_bridge/exporter.py:24-34); graph_summary is the tensor-free JSON the C loader consumes (exporter.py:86-99); save_graph writes the version-stamped spikeforge-nir-graph envelope v1 (spikeforge/nir_bridge/serialization.py:52-60). There is no reverse path: nothing loads C-produced weights back into a torch module or NIR graph.

Existing bridges (all Python → C)

• tools/export_fixture.py — generates the C test fixtures from to_nir + NirInterpreter (inference golden only).
• graph_summary JSON shape → nir_graph_load (include/nir_c_runtime/graph.h:88-97).
• Quantized-weight envelope (data/dtype/scale/zero_point) → dequantized at load (src/graph.c:32-76).

Sub-issues

• C Threshold is a strict Heaviside with no surrogate-gradient hook; no backward pass, loss, or optimizer exists #26 weight-update ops (surrogate-gradient hooks or STDP)
• No torch-free dataset loader: training data path is torchvision + DataLoader #27 torch-free dataset loader
• No seeded RNG or spike encoder in C: rate/latency/delta codings are torch-only #28 seeded RNG + spike encoder
• No checkpoint save/load in C: checkpoints are torch pickles and C JSON is parse-only #29 checkpoint save/load + JSON writer
• Reverse bridge missing: C-trained weights cannot flow back into NIR graphs or torch modules #30 reverse bridge (C weights back into NIR/torch) Impact if done: on-target fine-tuning / federated-edge learning without shipping Python; honest scope if not: the README "inference-only" line stays accurate.

Acceptance criteria (epic-level)

• Each sub-issue below is either implemented with C tests + a Python cross-check, or explicitly closed as out-of-scope with the README updated to say so.
• Any new C↔Python format is version-stamped like the existing spikeforge-nir-graph v1 envelope.