Derived _per_capita series are not rebuilt by reforms applied to a built system, and hold their value through a None total #9845

add_per_capita_uprating (policyengine_us/tools/per_capita_uprating.py) derives each <national total>_per_capita series from its national total and the population series. In two cases a derived series falls out of step with its total. Both were run on the head of #9628 (743b6dfd3d, which is main at ea79c39fb7 plus that PR), and neither involves backdating.

1. A dict reform applied to a built system does not rebuild them

CountryTaxBenefitSystem.modify_parameters rebuilds the derived series after init. But two other paths set values through Parameter.update directly, never call modify_parameters, and so never rebuild them:

• Reform.from_dict's apply (policyengine-core 3.32.15, reforms/reform.py);
• the YAML test runner's inline reforms, which assign self.parameters = modifier(self.parameters) (tools/test_runner.py).

from policyengine_core.reforms import Reform
from policyengine_us.system import CountryTaxBenefitSystem, system

TOTAL = "calibration.gov.cbo.income_by_source.employment_income"
reform = {TOTAL: {"year:2020:1": 1}}
built = CountryTaxBenefitSystem(reform=reform).parameters
afterwards = Reform.from_dict(reform, country_id="us")(system).parameters

The exhaustive reform in #9628's tests touches 42,489 parameters. Under it, 1 of the 19 derived series differs between the two paths: employment_income_per_capita in 2018 is 37,207.11 when built with the reform and 24,804.74 when the reform is applied afterwards.

2. A total set to None holds the previous per-capita value

_per_capita_values skips any instant where the total is None (if total_value is None or not population_value: continue). So the derived series keeps the previous value through that instant instead of becoming undefined:

Found while reviewing #9628. Its new whole-tree test skips the derived series because of case 1.