Fix connection generator code to work with PyNEST-NG

[jessica-mitchell]

This is a solution found by Claude code.
csa_example.py and csa_spatial_example.py both worked in Docker container with these modifications
The following text and code changes were done with AI:

File	Change
pynest/pynestkernel_aux.h	Remove conngenmodule.h include; replace Datum*-returning CYTHON_unpackConnectionGeneratorDatum with CYTHON_insertConnectionGenerator that stores shared_ptr<ConnectionGenerator> into a boost::any Dictionary slot
pynest/nestkernel_api.pxd	Add cdef extern declarations for CYTHON_isConnectionGenerator and CYTHON_insertConnectionGenerator
pynest/nestkernel_api.pyx	Add elif CYTHON_isConnectionGenerator branch in pydict_to_Dictionary before the AttributeError fallthrough
pynest/CMakeLists.txt	Add ${LIBNEUROSIM_LIBRARIES} to nestkernel_api link libraries
nestkernel/conn_builder_conngen.cpp	Read params_map indices as long not size_t; static_cast to size_t

Fix: ConnectionGenerator (libneurosim/CSA) support in pynest-ng

Background

When pynest-ng (Cython) was merged into main, the ConnectionGenerator
integration from the old SWIG-based pynest was not fully ported. There are
three distinct problems:

Problem 1 — pynestkernel_aux.h exists but is broken and unused

The file pynest/pynestkernel_aux.h was carried over from the SWIG era but
was never wired into pynest-ng. It has two bugs that make it unusable as-is:

Bug A — includes a module that no longer exists:

// Includes from conngen:
#include "conngenmodule.h"    // <-- ConngenModule was removed from NEST main

This line causes a compile error if the header is ever included while
HAVE_LIBNEUROSIM is defined.

Bug B — returns Datum*, which is incompatible with pynest-ng:

Datum*
CYTHON_unpackConnectionGeneratorDatum( PyObject* obj )
{
  ...
  ret = static_cast< Datum* >( new nest::ConnectionGeneratorDatum( cg ) );
  return ret;
}

pynest-ng does not use the SLI Datum type. All values are stored in
boost::any fields inside a Dictionary. A function returning Datum* cannot
be used to insert a ConnectionGenerator into a pynest-ng Dictionary.

Problem 2 — pydict_to_Dictionary has no ConnectionGenerator branch

nestkernel_api.pyx falls through to raise AttributeError for any type it
does not recognise. pynestkernel_aux.h is not included and its macros are
never called:

AttributeError: value of key (cg) is not a known type,
got <class 'csa.connset.ConnectionSet'>

Problem 3 — conn_builder_conngen.cpp reads params_map as size_t

boost::any requires an exact type match. pydict_to_Dictionary stores Python
int values as long, but the conngen builder reads them as size_t:

NESTErrors.TypeMismatch: Failed to cast 'delay' from long to type unsigned long

---

Reproducer

With NEST built with libneurosim support:

import nest
import csa

pre  = nest.Create("iaf_psc_alpha", 4)
post = nest.Create("iaf_psc_alpha", 4)

cg = csa.cset(csa.oneToOne, 10000.0, 2.0)
params_map = {"weight": 0, "delay": 1}
connspec = {"rule": "conngen", "cg": cg, "params_map": params_map}

nest.Connect(pre, post, connspec)  # raises AttributeError  (problem 2)
                                   # after fixing problem 2: raises TypeMismatch (problem 3)

The existing test suite in
testsuite/pytests/connect/test_connect_conngen.py covers all four conngen
scenarios and serves as the regression suite for this fix.

[heplesser]

Solving this will be more complicated, since we need to figure out how to interface ConnectionGenerator from libneurosim with NEST. That should only happen once we have #3748 merged, so we do not first create a solution based on boost::any which we then immediately need to convert to variant.

It turns out that our CSA tests have not run for a long time in the CI, they are always skipped because either CSA or libneurosim is not properly installed.

[jessica-mitchell]

OK should i Just close this one then?

[heplesser]

OK should i Just close this one then?

Leave it open, it contains some useful information.