subsample() leaves the baseline branch on the reform system and strands the baseline system under branches["tax_benefit_system"]

[MaxGhenis]

Summary

After Simulation.subsample() on a reform simulation, the rebuilt baseline branch carries the reform policy, and the saved baseline tax-benefit system is stranded under a stray branches["tax_benefit_system"] dict key. Anything that reads the baseline branch after subsampling (e.g. labor-supply behavioral responses, which compare against simulation.get_branch("baseline")) silently compares reform against reform.

Where

simulation.py, end of subsample() — master L1854–1858:

# Ensure the baseline branch has the new data.
if "baseline" in self.branches:
    baseline_tax_benefit_system = self.branches["baseline"].tax_benefit_system
    self.branches["baseline"] = self.clone()
    self.branches["tax_benefit_system"] = baseline_tax_benefit_system

self.clone() clones the reform simulation (including a clone of its reformed system), and the last line assigns the saved baseline system into the branches dict under a string key instead of restoring it onto the rebuilt branch (self.branches["baseline"].tax_benefit_system = ...). The rebuilt branch also never gets branch_name = "baseline" / parent_branch set the way get_branch would set them.

Executed repro (country template)

from policyengine_core.country_template import Microsimulation

REFORM_RATE = 0.42
INSTANT = "2022-01-01"

simulation = Microsimulation(reform={"taxes.income_tax_rate": {INSTANT: REFORM_RATE}})
baseline_system_before = simulation.branches["baseline"].tax_benefit_system
# IDs are ETERNITY inputs while weights are annual, so pass the weight period.
simulation.subsample(n=1, seed="phase-d", time_period="2022")

baseline_system_after = simulation.branches["baseline"].tax_benefit_system
print(list(simulation.branches))
print(baseline_system_after.parameters.taxes.income_tax_rate(INSTANT))
print("tax_benefit_system" in simulation.branches)
print(simulation.branches["tax_benefit_system"] is baseline_system_before)

Output (master):

['baseline', 'tax_benefit_system']
0.42        # baseline branch now carries the reform rate; true baseline is 0.15
True        # stray dict key exists
True        # ...and holds the original baseline system

Expected

After subsampling, branches["baseline"] should hold the subsampled populations with the original baseline tax-benefit system (rate 0.15 here), and no "tax_benefit_system" key should appear in branches.

Context

• The block came from the fix for Subsample doesn't work with behavioural response simulations #296 (subsample vs behavioral-response array sizes).
• Found while auditing the branch machinery for Substitution channel of behavioral responses is inert: LSR measurement branches record MTR = 1.0 for every adult policyengine-us#9086; the delete-arrays fix in Purge the full visibility chain in Simulation.delete_arrays and give clones private disk views #521 does not touch this (its audit confirmed subsample() calls neither delete API).
• Repro was executed on master (country template, exit 0, asserts pass).

🤖 Generated with Claude Code