[juaristi22]

Fixes #2297

Default input exports dropped dataset-supplied values for variables with formulas. In the reported example, £9,500 of State Pension became £0 after export and reload. Export surviving supplied arrays, including formula and adds inputs, while excluding calculated caches and model-generated values. Preserve period/entity handling, identifiers and the existing include_computed_variables=True path.

Reproducer: policyengine_uk/tests/test_input_dataframe_export.py::test_dataset_formula_input_round_trip_preserves_final_pension

Validation: 51 distinct export, DataFrame and input-reset cases pass independently on each Core version: installed 3.32.27 and an official-wheel 3.33.2 package overlay. The pension regression reproduces £9,500 → £0 before the fix and £9,500 → £9,500 afterward. Native formatting/lint passes; independent code and engineering/source reviews completed with no findings or material gaps. Runtime: Python 3.13.14 and microdf 1.2.1; full-lock environment validation was not run.

axiom: n/a — input serialization/provenance engineering; policy formulas and parameters are unchanged. No encoded-parity claim.

[vahid-ahmadi]

Looks right, fine to approve.

@juaristi22, the default export now keeps supplied values for variables with formulas. The state_pension_reported round trip in #2297 keeps £9,500, and include_computed_variables=True still goes through Core unchanged. Core 3.33.2's own to_input_dataframe excludes formula variables by design, so overriding it in UK is the right place for the fix.

1. should: policyengine_uk/utils/input_dataframe_export.py:8-31, 39, 62 adds new reads of Core private state: holder._user_input_storage, _memory_storage, _disk_storage, _stores_user_input, simulation._get_visible_branch_names and, when present, simulation._get_recorded_input_array. #2175 (and #2495 for the parameter side) is moving the UK package off exactly these fields. Please either add this module to #2175's checklist, or raise the Core floor to 3.33.2 and drop _legacy_recorded_array. Otherwise the fallback will copy Core internals that can change underneath it.
2. nit: _legacy_recorded_array mirrors Core 3.33.2's tier precedence by hand (memory then disk, skipping derived values). A test that runs both paths on the same fixture and asserts the same arrays would catch drift. If point 1 drops the fallback, this goes too.

What I checked

• Read the full diff and issue #2297, and compared the code with Core 3.33.2's _get_recorded_input_array and to_input_dataframe.
• pytest test_input_dataframe_export.py test_dataframe_dataset.py: 103 passed (locked Core 3.33.2).
• Reverted simulation.py to main. test_input_dataframe_export.py then fails 14 of 25.
• ruff check and ruff format --check on the changed Python files: clean.
• Hosted CI on this head: all green.
• Not checked: the Core 3.32.27 fallback path, since my environment has only the locked 3.33.2.

🤖 Generated with Claude Code

[vahid-ahmadi]

Approving, per my review above. The shoulds and nits can follow.