Model.set_param does not rebuild a species initial condition that references the parameter (species_ic_param_refs recorded, ic_state_dirty never set)

[wshlavacek]

Model.set_param updates the parameter but leaves any species initial condition that references it at its network-generation value. get_state() returns the stale amount, and reset() does not rebuild it either.

The dependency is already recorded and the invalidation flag already exists — NetworkModel.species_ic_param_refs names the (species, parameter) pair, and NetworkModel.ic_state_dirty stays False across the set_param call. So the machinery to invalidate appears to be in place and simply does not fire.

This is the network-path counterpart of #44, which fixed the same class of defect (a set_param that does not re-derive seed-species amounts) on the network-free path.

Reproducer

Self-contained, two species:

cat > minimal.net <<'EOF'
begin parameters
    1 Stot  100
    2 k     1
end parameters
begin species
    1 A()  Stot
    2 B()  0
end species
begin reactions
    1 1 2 k #A->B
end reactions
begin groups
    1 Obs_A  1
end groups
EOF

import bngsim

m = bngsim.Model.from_net("minimal.net")

print(m._core.species_ic_param_refs)  # [(0, 0)]  species 0's IC references parameter 0
print(m._core.ic_state_dirty)         # False
print(m.get_state()[0])               # 100.0

m.set_param("Stot", 1e6)

print(m.get_param("Stot"))            # 1000000.0   <- parameter did move
print(m._core.ic_state_dirty)         # False       <- never marked dirty
print(m.get_state()[0])               # 100.0       <- expected 1000000.0

m.reset()
print(m.get_state()[0])               # 100.0       <- reset does not rebuild it either

	expected	actual
get_param("Stot")	1e6	1e6 ✓
get_state()[0]	1e6	100.0 ✗
get_state()[0] after reset()	1e6	100.0 ✗
ic_state_dirty after set_param	True	False ✗

On a published model

Same behaviour on a real network — scaffolded_mapk_cascade_kocieniewski2012 (85 species), where Stot is parameter 5 and species 3's initial condition (Scaff(map2k,map3k,mapk) Stot), and species_ic_param_refs reads [(3, 5)]:

m = bngsim.Model.from_net("B02_Kocieniewski_2012.net")
m.get_state()[3]                 # 100000.0
m.set_param("Stot", 1e6)
m.get_state()[3]                 # 100000.0   <- expected 1e6

Why it matters

A dose scan over a total amount — the common case, since a "total" is exactly the kind of parameter a species initial condition is written in terms of — silently returns identical results at every dose, with no error and no warning. Nothing in the API surface indicates the scan did not take effect: get_param confirms the new value.

It is also a cross-engine trap. AMICI, reading the same system through SBML initialAssignment, moves correctly, so a comparison scanning such a parameter disagrees for a reason that has nothing to do with the solver under test. Found while building a cross-engine steady-state dose scan, where it affected 3 of 6 published models (Stot, Rtot, APCtot) and had to be worked around by rewriting the parameter in the .net and reloading.

Version

bngsim 0.11.35 (build v0.11.35-72-g201836f), macOS 15 / darwin 24.6.0, Python 3.12.

Suggested fix

Have set_param / set_params consult species_ic_param_refs and set ic_state_dirty when the written parameter appears in it, so the existing rebuild path runs. Worth deciding explicitly whether the IC rebuild happens eagerly on set_param or lazily at the next get_state()/reset()/solve; the lazy form matches what the dirty flag implies. reset() in particular should rebuild from current parameter values rather than restore a snapshot taken at load time.