feat(study): reject parameter dependency extension at load time, with actionable errors

[aoustry]

Process ID

Process: GP-02 (refs #258 — see Scope vs #258 below; this PR does not close it)

Description

A component parameter's time/scenario dependency must be ≤ the one its model declares, axis by axis: a component may narrow what the model declares (true → false) but never extend it.

Narrowing already worked, end to end — this PR adds the missing test coverage for it, it does not fix it. Extension already failed too; what was wrong is when and how. It surfaced only inside build_problem() as Data inconsistency for component: G, parameter: cost. Requirement not met., which named neither the flags, nor the model, nor the file — so load_study() succeeded and nothing failed until you tried to solve.

• The check moves into resolve_system(), so load_study() and the CLI fail immediately with an error naming the component, the parameter, and both declarations.
• Study.check_consistency() stays as the safety net for programmatically built studies, now reporting the data's and the model's axes.
• Data-shape errors name the offending component/parameter, report the shape read, and point at the flag to change. Strict shapes are unchanged (T×1, 1×S, T×S, inline float).
• check_requirement() is now derived from a new indexing_structure() per data container instead of being reimplemented four times; accept/reject behaviour is identical, and four dead isinstance(self, X) guards (tautologically true inside X.check_requirement) are gone.

So the user-visible delta is error timing and wording only — no study that used to work stops working, and no study that used to fail starts working.

Scope vs #258

#258 reports three things. This PR addresses one and takes a position on a second; the third is untouched. It therefore refs the issue rather than closing it.

#258 asks	This PR
Narrowing (true → false from the system file) should work	Already worked before this PR — covered by tests here, no behaviour change
The system file should be able to extend the model's declared dependency	Rejected by design, earlier and more clearly — the opposite of the request
A single-axis parameter shouldn't be capped at one column/row of data	Not addressed; only the error message is reworded

Rationale for rejecting extension: _build_param_arrays_for_model builds one array per model, shaped from param.structure (the model's declaration). A component varying along an axis its model does not declare has nowhere to live without per-component dims, which would ripple into the indexing-structure analysis used for constraint shapes. Making it a clear, early error is the cheap and coherent option.

Impact Analysis

Affected modules: study/ only (resolve_components.py, data.py, study.py). No changes to expression/, simulation/, or optim_config/.

Solver output values are not expected to change. Parameter arrays keep the dimensions the model declares, and narrower data is broadcast over the remaining axes exactly as before; the check_requirement truth table is unchanged for all four data containers. test_narrowing_does_not_change_the_problem_shape asserts narrowing leaves array dimensions, variables, and constraints untouched, and the 8 parametrized end-to-end cases pin the objective values. Backward compatible; the only wording change to an existing message is scenario-series "exactly one row" instead of "exactly one line".

One caveat worth a reviewer's eye: moving the check into resolve_system() means a study whose system file extends a model declaration now fails at load, not at solve. That is the intent, but it is a behaviour change for gems_craft-only consumers that load without solving. gems_craft_hybrid is unaffected (it parses and writes, it never calls resolve_system).

AbstractDataStructure gains a new abstract indexing_structure(), so any out-of-tree subclass must implement it.

Checklist

• Unit tests pass (pytest) — 632 passed, 6 skipped, 2 xfailed
• Type checking passes (mypy) — 68 source files
• Formatting passes (black, isort)
• pyproject.toml version bumped if applicable — not bumped: this repo bumps in dedicated release: commits, so the entry went under [Unreleased] in docs/CHANGELOG.md
• AGENTS.md reviewed for impact and updated if needed — reviewed, no change needed (the rule is user-facing and documented in docs/user-guide/inputs.md)