No validation that Variable lower_bound/upper_bound indexing is compatible with the variable's own IndexingStructure

[tbittar]

Summary

Variable.lower_bound / Variable.upper_bound expressions are only validated for being constant with respect to decision variables (is_constant, i.e. compute_degree(expr) == 0 in variable.py). There is no check that the bound expression's own time/scenario indexing is compatible with (a subset of) the variable's structure: IndexingStructure.

Details

• A bound expression can reference parameters that are themselves time_dependent/scenario_dependent (resolve_library.py:408-413), regardless of whether the variable itself declares time_dependent/scenario_dependent.
• Model.__post_init__ and _make_structure_provider in model.py do cross-check indexing for objective contributions — they require compute_indexation(contribution, provider) == IndexingStructure(False, False) (model.py:139) — but no equivalent check exists for Variable.lower_bound / Variable.upper_bound against Variable.structure.
• The mismatch only surfaces downstream, at simulation-build time, in _add_variable_for_group / _to_bound_array (optimization.py:794-985). The bound expression is vectorized against the variable's var_shape/dims (derived from var.structure), and _to_bound_array only knows how to add missing axes for dims the variable has but the bound value lacks. It has no handling for the reverse case: a bound whose underlying parameter carries a time or scenario axis that the variable's own dims don't include.

Impact

When a bound's effective indexing exceeds the variable's own indexing:

• Usually: a numpy broadcast shape-mismatch error at solve-build time, with no context pointing back to the offending variable/bound definition.
• Worst case: if the extra axis happens to match another axis's length (e.g. number of time steps equals number of components), np.broadcast_to silently succeeds and produces silently wrong bound values — no error at all.

Suggested fix

Add an explicit check (either in Variable.__post_init__/Model.__post_init__, or wherever the structure provider is available) that compute_indexation(bound_expr, provider) is a "subset" of var.structure for both lower_bound and upper_bound, and raise a clear ValueError at model-definition/resolution time instead of letting it fail (or silently misbehave) deep in optimization.py.

Where this came from

Surfaced while reviewing Variable (variable.py:23-33) on feat/lb_ub_operator.