Keep the inputs a reform applied after the build stores through reset_calculations #617

A reform applied to a simulation after it is built (a Reform passed as reform=, or apply_reform on a built simulation, a clone or a branch) can store inputs with set_input. input_variables is taken when the simulation is built, and reset_calculations deletes the values of every variable it does not name, so a reset deletes what such a reform stored and recalculates the formula instead.

PolicyEngine/policyengine-uk#2075 works around this in policyengine_uk/simulation.py: while Simulation.apply_reform runs, _user_input_keys is replaced by a set subclass that logs each (variable, branch, period) key core adds, and afterwards apply_reform adds to input_variables every logged variable the simulation still holds under a branch it can see (_get_visible_branch_names). That relies on core internals: the name and shape of _user_input_keys, the fact that Holder._set adds a key on every storing route, and _get_visible_branch_names. The country package's tests fail if core changes them (the reform's inputs stop surviving a reset), but the logic belongs in core.

Proposal: core's Simulation.apply_reform records the input keys a reform stores and extends input_variables with them (with a new list, since a clone or branch shares its parent's), so every country package gets this, and policyengine-uk can drop its override. policyengine-uk's test_reform_inputs_survive_reset.py and test_a_reform_applied_later_keeps_what_it_sets_through_a_reset describe the expected behaviour, including inputs a reform replaces, values core does not store, nested reforms and a reform that raises.

Raised in Vahid Ahmadi's review of PolicyEngine/policyengine-uk#2075.

🤖 Generated with Claude Code