A parameter fixed by an SBML initialAssignment goes stale on the fast path, so a fitted dependency simulates the wrong model

[wshlavacek]

Summary

On the sbml_backend = bngsim fast path (the cached-clone + set_param route added in #415), a
parameter whose value is fixed by an initialAssignment is never recomputed. bngsim evaluates
the assignment once when the engine template is loaded, and set_param does not propagate to it, so
the derived parameter keeps its load-time value for the whole fit even when the parameters it is
derived from are fitted or set by a condition:.

_compute_initial_dependency_names only ever seeded its dependency set from initialAssignments on
species; an assignment on a parameter was collected into expr_refs but never marked as needing
a recompute.

Impact

The simulated model silently disagrees with what the same job produces through the reload path (and
with what any other PEtab tool produces). It is a scalar-path bug: it corrupts the objective for
every optimizer, not just the gradient ones.

Bertozzi_PNAS2020 (Benchmark-Models-PEtab, in the Grein et al. 2026 subset) is the worked example:

<initialAssignment symbol="beta_N">   <!-- beta_N = R0_ * gamma_ / N_ -->

beta_N is the only place R0_ reaches the dynamics, and the conditions set
R0_ = R0_CA / R0_ = R0_NY and N_ = 39560000. Observed before the fix:

base I_   : [5.00e+02 1.46e+07 5.35e+06 1.97e+06]
R0_ 3 -> 3.3    max trajectory delta = 0        # R0_ is completely inert
engine beta_N (fast path)  = 0.01               # the load-time value (R0_=0.1, gamma_=0.1, N_=1)
engine beta_N (reload)     = 8.34e-09           # the correct R0_*gamma_/N_

Two of the eight free parameters (R0_CA, R0_NY) therefore had an identically flat objective, and
the reported nominal NLL was 1.79e11 against a reference J* = 158.86. With the fix the fast path
is byte-identical to the reload path and the trajectory tracks the data.

Fix

Treat a parameter with an initialAssignment as the derived constant SBML says it is: seed the
initial-dependency set from parameter assignments as well as species ones, recompute both in
_recompute_initial_assignments (via libSBML's expandInitialAssignments, as the species initials
already are), and apply the recomputed parameter values to the engine clone alongside the
recomputed species initials. A parameter an assignmentRule or rateRule governs is excluded --
bngsim recomputes those dynamically, so an initial override would fight the rule.

Scope

bngsim SBML/Antimony backend only. The reload path (_needs_structural_reload) was already correct
and is the oracle the regression test compares against.