Port the BNGL parameter-expression evaluator to libpetab-python's BnglModel, then delete the staging copy

[wshlavacek]

Context: #591 retired PyBNF's local BnglModel adapter once petab 0.9.0 shipped the native language: bngl loader (PEtab-dev/libpetab-python#508). That adapter was the only consumer of pybnf/petab/_bngl_expr.py, the BNGL parameter-expression evaluator from #666 (PR #673), whose module docstring says it is stdlib-only precisely so it can travel to libpetab-python.

The gap

petab's native BnglModel does not evaluate an expression-valued parameter:

• get_parameter_value raises NotImplementedError ("Evaluating a BNGL parameter expression requires BNG2.pl / network generation, which is out of scope for the introspection-only BnglModel").
• get_free_parameter_ids_with_values catches the ValueError and silently skips the parameter.

That is exactly the pre-#666 behaviour, now on the petab-side validation path: a language: bngl problem whose model defines a parameter as an expression over other parameters (20.8% of declarations across the corpora on hand, per #673) loses that parameter from petab's parameter-table checks with nothing said.

This is not a regression introduced by #591. On petab 0.9.0 register_bngl() was already a no-op, so every 0.9.0 installation was already validating through the native class. But the #666 fix never reached the path it was written for.

What to do

Port the evaluator upstream: a PR to PEtab-dev/libpetab-python that gives BnglModel.get_parameter_value / get_free_parameter_ids_with_values the dependency-ordered, BNG2.pl-verified evaluation _bngl_expr implements (its semantics are pinned against BNG2.pl 2.9.3 in tests/test_petab_bngl_expr.py, which should travel too). Once a petab release carries it, delete pybnf/petab/_bngl_expr.py and its tests here and bump the petab floor.

Until then _bngl_expr.py stays in the tree as the staging copy, with no runtime consumer.

[wshlavacek]

Correction to the description above, after tracing every caller of the two methods in petab 0.9.0 and in PyBNF: this gap does not affect PyBNF, and it does not affect petab's validator either.

• PyBNF calls neither method. The importer reads the model through PyBNF's own reader and uses only parameter names (for the observable-formula namespace and the shadow check). The exporter takes nominal values from the conf's free and fixed parameters and carries the model file verbatim, so BNG2.pl evaluates any expression at simulation time. Neither path ever asked the adapter for a value, before or after Evaluate BNGL parameter expressions instead of dropping them #673.
• petab v2 lint calls neither method. No lint task touches get_parameter_value or get_free_parameter_ids_with_values, so petablint on a BNGL problem with expression-valued parameters is unaffected. The sentence in the description about losing the parameter "from petab's parameter-table checks" is wrong.
• Where it does bite: two petab v1 helpers. petab.v1.parameters.create_parameter_df, which fills a parameter table's nominal values from the model, and petab.v1.parameter_mapping, which simulators such as AMICI or pyPESTO use to bind model parameters per condition. In both, an expression-valued BNGL parameter is silently skipped.

So the port is an improvement for petab users of BNGL models and for any future simulator that consumes a BNGL PEtab problem through petab's parameter mapping, not something PyBNF is waiting on. A PyBNF-side fix would mean re-registering a subclass into petab's model factory, which is the register_bngl() shim #591 just retired, and it would help only code that imports PyBNF first. Wrong layer.

Decision: keep this open as low priority. The evaluator stays in the tree as the staging copy for the upstream port; if the port is ever dropped instead, delete pybnf/petab/_bngl_expr.py and tests/test_petab_bngl_expr.py and close this.