parse_bngl_protocol drops saveParameters/resetParameters silently, so a ProtocolSpec keeps a setParameter that BNG2.pl reset #1478

What happens

parse_bngl_protocol, and with it Model.from_bngl(..., protocol=True), lists saveParameters and resetParameters among the pure build/IO directives (_BUILD_VERBS, python/bngsim/convert/_protocol.py:110-111). The comment there says none of them "alters the trajectory a consumer would compute". So both are recorded in spec.dropped and otherwise ignored: no warning, nothing in spec.lossy, and strict=True accepts them too.

But resetParameters() puts every constant parameter back to the values saveParameters() saved (BioNetGen 2.9.3, Perl2/BNGModel.pm, resetParameters). A setParameter between the two is undone before the next run, and the spec keeps it.

Repro

begin model
begin parameters
  k 1
end parameters
begin molecule types
  A(b)
end molecule types
begin seed species
  A(b) 100
end seed species
begin observables
  Molecules Atot A()
end observables
begin reaction rules
  A(b) -> 0 k
end reaction rules
end model
generate_network({overwrite=>1})
saveParameters()
setParameter("k",2)
resetParameters()
simulate({method=>"ode",t_end=>1,n_steps=>1})

spec = bngsim.convert.parse_bngl_protocol("m.bngl", strict=False)
spec.steps    # set_parameter k=2.0, then the simulate
spec.dropped  # ('generate_network', 'saveParameters', 'resetParameters')
spec.lossy    # ()

No warning is raised, with or without strict. BNG2.pl, run on the same file, prints Reloaded parameters saved with label 'DEFAULT', writes k 1 into the .net it runs, and reports Atot 36.79 at t=1 (100·e⁻¹). Replaying the spec runs with k=2, which gives 13.53 (100·e⁻²).

In the corpus

18 of the 1,070 tracked .bngl files use saveParameters/resetParameters. In 4 of them, a setParameter falls between a save and a reset that comes before a later run, so the spec is wrong for that run:

• parity_checks/bng_parity/models/slow/rulehub/Published/Lin2019/TCR_model.bngl
• parity_checks/bng_parity/models/glacial/rulehub/Published/Lin2019/ERK_model.bngl
• parity_checks/bng_parity/models/glacial/rulehub/Published/Lin2019/prion_model.bngl
• parity_checks/bng_parity/models/slow/rulehub/Published/Salazar-Cavazos2019/190127_CHO_EGFR_best-fit.bngl

In TCR_model.bngl, setParameter("lambda",0.1) precedes the std_scaling run, and resetParameters() follows it. The spec runs partial_scaling with lambda 0.1, where BNG2.pl runs it with the declared 1.

None of the 14 models #1366 lists uses either verb, so this does not affect that issue's models.

Who reads the spec

• Model.from_bngl(protocol=True). Since bngl: warn when a source's first run starts from a state from_bngl drops (#1366) #1388, its docstring says the spec returns the dropped actions instead of the dropped-state warning.
• The SED-ML/OMEX conversion in bngsim.convert, which parses with strict=strict.
• parity_checks/codegen_paths/_metrics.py, through primary_experiment().

Fix directions

1. Carry them. Add save_parameters/reset_parameters StateChange kinds with labels, as save_concentrations/reset_concentrations already have (issue Named saved concentration states + reset_conc-to-snapshot scan (parameter_scan/bifurcate after pre-equilibration) #11), and have the consumers apply them.
2. At least, flag them. Treat them as lossy, so a non-strict parse warns and records them in spec.lossy, and strict=True refuses. Refusing only when a setParameter falls between a save and its reset would leave alone the 14 files where the pair changes nothing.

Related: addConcentration and bifurcate are not carried either, but the parser does warn about them ("unrecognized BNGL action").

Found while reviewing #1388 (issue #1366).