poc: add BrianExporter and BrianImporter skeleton

[Sanchit2662]

What this is

I'm applying for GSoC 2026 to work on serialization/deserialization for Brian2 models, results, and input data. This is a proof-of-concept showing I understand the existing brian2tools codebase and have a concrete approach to the problem.

This is the brian2tools half of the PoC. The other half lives in the brian2 repo and demonstrates portable state checkpointing using _full_state()/_restore_from_full_state(): brian-team/brian2#1799

---

What I built

brian2tools/baseexport/exporter.py
BrianExporter.serialize(net, 'file.brian')

I built this by wrapping the existing collect_*() functions from collector.py using the same COLLECTOR_MAP pattern I found in BaseExporter.network_run() (device.py:151). While reading through the code I identified three gaps that BaseExporter deliberately omits because it operates before the simulation runs:

• _json_safe() : converts Quantity objects in collector output to JSON-safe dicts. The concrete problem: collect_Equations() (collector.py:212) stores eqs.unit as a raw Quantity, which breaks json.dumps(). I serialize it as {'__type__': 'quantity', 'value': ..., 'dim': [7-element SI exponent tuple]} using Dimension._dims directly.

• _collect_state() : captures actual state variable values via var.get_value() after the run. Since BaseExporter intercepts code generation before the simulation, it never sees the values.

• _collect_connectivity() : captures _synaptic_pre / _synaptic_post arrays. I noticed that collect_Synapses() (collector.py:565) only stores the connect() arguments condition string, p, n — not the resulting index arrays. For probabilistic connections (p=0.1), re-running that condition on load gives different synapses every time. Storing the arrays fixes this.

brian2tools/baseimport/importer.py
BrianImporter.load('file.brian') → (net, namespace)

I mapped each field in the collect_*() output back to the corresponding Brian2 constructor. For example, _reconstruct_neurongroup() pulls threshold/reset/refractory from ng_dict['events']['spike'] (the structure from collect_Events() at collector.py:225), and _reconstruct_synapses() resolves source/target using collect_SpikeSource()'s Subgroup dict form. Connectivity is restored via syn.connect(i=arr, j=arr) — not by re-running the original condition. State variables are restored after connect() so DynamicArrayVariable sizes are correct first.

11 tests in tests/test_poc_exporter.py covering archive structure, Quantity serialization, state capture, connectivity arrays, and full NeuronGroup/Synapses round-trips.

---

What this isn't

• Not a full implementation monitors, PoissonGroup, SpikeGeneratorGroup, SpatialNeuron reconstruction are planned but out of scope here
• Not a device-mode integration, I call it explicitly after net.run(), not via set_device()
• Not production-ready

---

Questions for reviewers

1. I'm using var.owner.name != obj.name to filter out variables owned by other objects (clocks etc.) — is there a more idiomatic way to do this check?

2. collect_Synapses() at collector.py:629 drops non-scalar delays. Should I capture them the same way I capture _synaptic_pre / _synaptic_post, or is there a better hook?

3. For Synapses reconstruction I use str(obj.equations) as the model string. Is there a more reliable canonical form I should use instead?