US reforms silently fail when applied via p.update() after Microsimulation construction

[anth-volk]

Summary

Reforms applied to US simulations via p.update() after Microsimulation construction have no effect on calculations. The parameter values are modified, but the tax calculations return baseline results, causing silent failures.

Root Cause

The US country package (policyengine-us) uses a shared singleton TaxBenefitSystem:

# policyengine_us/system.py
system = CountryTaxBenefitSystem()  # Module-level singleton

class Microsimulation(CoreMicrosimulation):
    default_tax_benefit_system_instance = system  # All sims share this

When p.update() is called after construction:

1. The parameter value is modified on the shared TaxBenefitSystem ✓
2. But calculations are cached at the simulation level with the old values
3. The updated parameter is never used in calculations ✗

Expected vs Actual

Metric	Expected	Actual
Parameter value	Changed ✓	Changed ✓
Tax calculation	Different	Unchanged ✗

Workaround

Pass the reform at Microsimulation construction time:

reform_dict = {
    "gov.irs.deductions.standard.amount.SINGLE": {
        "2024-01-01": 29200
    }
}
microsim = Microsimulation(reform=reform_dict)  # Works correctly

Comparison with UK

The UK package (policyengine-uk >= 2.72.3) was refactored in July 2025 to give each simulation its own TaxBenefitSystem instance, so p.update() works correctly there.

Package	Architecture	p.update() Works?
policyengine-uk	Separate TBS per simulation	✅ Yes
policyengine-us	Shared singleton TBS	❌ No

Impact

Any code using policyengine.py simulation_modifier pattern (which calls p.update()) for US simulations will silently return baseline results regardless of the reform.

Suggested Fix

Either:

1. In policyengine.py: Convert parameter values to reform dict and pass at construction (implemented in this repo)
2. In policyengine-us: Refactor to match UK architecture (separate TBS per simulation)