Measurement-formula namespace advertises SBML assignment-rule variables that aren't resolvable at materialize → fit fails mid-run #464

Summary

A measurement-model observable: <id>, formula: <expr> (ADR-0036) that references an SBML assignment-rule variable passes config-time validation but then fails at runtime, mid-fit, because assignment-rule variables are in the formula-validation namespace yet are never simulation-output columns. The validation namespace and the set of symbols actually resolvable at materialize are out of sync: validation says yes, the run says no.

This is distinct from #463 (which is about .ant models routing through parse_bngl). This one affects any SBML model (.xml and .ant), fails late (during the fit) rather than at config build, and is the opposite timing/path. Surfaced while finishing #386 deliverable 6 / #462.

Mechanism

1. Configuration._model_expression_namespace (pybnf/config.py:1858) builds the allowed-symbol set for an SBML model from pybnf.petab._sbml.parse_model(...).namespace_symbols, which includes assignment-rule targets (e.g. the Becker model's Epo_medium := Epo + dEpoe, Epo_cells := Epo_EpoRi + dEpoi).
2. _load_measurement_models (pybnf/config.py:1764) validates each observableFormula's free symbols against that set — so a formula referencing an assignment-rule variable passes and a MeasurementModel is built.
3. At fit time, MeasurementModel.materialize (pybnf/measurement/base.py:101) resolves each symbol to a simulation-output column, a PSet value, or a constant. bngsim's SBML simulation outputs species only (not assignment-rule variables), and the symbol is neither a free parameter nor a constant — so it raises:

PybnfError: Measurement model '<id>' references 'Epo_cells', which is neither a
simulation-output column (...) nor a fit/model parameter. The measurement model
cannot be evaluated.

Reproduction

With tests/sbml_files/becker_epor.xml (sbml_backend = bngsim):

edition = 2
job_type = trf
objective = chi_sq
sbml_backend = bngsim
model: becker_epor.xml
observable: Epo_cells, formula: Epo_cells     # Epo_cells is an SBML assignment rule
experiment: tc, data: tc.exp
uniform_var = kon 1e-5 1e-2
uniform_var = koff 1e-3 1e-1

• parse_sbml(...).namespace_symbols contains Epo_cells → True.
• Configuration(...) builds successfully (the measurement model Epo_cells is created).
• Running the fit raises the materialize error above (the bngsim SBML output columns are time, EpoR, Epo, Epo_EpoR, Epo_EpoRi, dEpoi, dEpoe — no Epo_cells / Epo_medium).

The working alternative is to reconstruct the observable from species — formula: Epo_EpoRi + dEpoi — which is the intended observation-layer path and what tests/test_gradient_optimizer.py::test_trf_multistart_smoke_on_becker_epor_sbml_measurement_layer (#462) does. The bug is specifically that referencing the assignment rule directly is accepted and then fails late.

Why it matters

• Config validation is supposed to be the typo-guard ("an unknown symbol is an error, not a new free parameter") but here it green-lights a symbol the run can't resolve, so the error lands deep in a fit instead of at load.
• It's a natural thing to write: SBML authors define exactly these convenience variables (the D2D Epo_cells / Epo_medium observables are assignment rules), so reaching for them by name is the obvious first attempt.

Suggested fix

Reconcile the validation namespace with what's resolvable at materialize. Either:

1. Exclude un-resolvable assignment-rule variables from namespace_symbols (or have _load_measurement_models validate against species ∪ parameters ∪ compartments ∪ free-params only), so referencing one fails at config build with the existing pointed message; or
2. Make assignment-rule variables resolvable — output them as columns (roadrunner already exposes them), or evaluate the rule in the measurement layer — so formula: Epo_cells simply works.

Option 1 is the smaller, safer fix (fail fast, clear message); option 2 is the more user-friendly one. Either way, add a test for an assignment-rule symbol in a formula.

References

• pybnf/config.py:1858 (_model_expression_namespace), pybnf/config.py:1764 (_load_measurement_models), pybnf/measurement/base.py:101 (MeasurementModel.materialize).
• Sibling issue Measurement-model formulas silently unavailable on Antimony (.ant) models — namespace built from parse_bngl #463 (the .ant → parse_bngl namespace gap). Parent context: Gradient-based local optimizers (D2D-style: TRF/Levenberg-Marquardt + L-BFGS-B) in the async Algorithm loop #386 deliverable 6 / Gradient optimizers: finish #386 deliverable 6 — convergence-vs-baseline test + becker smoke test #462.