Jointly calibrate ACA take-up to CMS APTC people and SOI PTC returns

[daphnehanse11]

Context

The current published Populace US 2024 dataset no longer has degenerate ACA inputs: takes_up_aca_if_eligible is nonconstant and selected_marketplace_plan_benchmark_ratio has real variation. The remaining ACA mismatch looks like take-up composition, not a selected amount-variable bug.

Using sim.calc on populace.data.load("us", 2024):

• CMS APTC-recipient person count: simulated 17.86M vs target 19.74M (-9.6%).
• SOI PTC returns: simulated 8.39M vs target 7.84M (+6.9%).
• SOI PTC amount: simulated assigned/used PTC $56.82B vs target $53.91B (+5.4%).
• used_aca_ptc and assigned_aca_ptc are effectively identical, so the issue is not a wrong comparison between assigned and used PTC.
• The implied composition is off: 17.86M / 8.39M = 2.13 APTC people per positive PTC tax unit, compared with 19.74M / 7.84M = 2.52 from CMS/SOI targets.

This is related to umbrella #70, but narrower than #74/#75/#76 and not covered by active PRs #165, #166, or #167.

Goal

Improve ACA source-stage take-up assignment so it jointly respects CMS APTC person counts and SOI PTC return/amount targets, rather than matching one surface while creating a contradictory composition on the other.

Suggested implementation

• Add an ACA diagnostic that reports:
  • APTC-recipient people, using the current target semantics: assigned_aca_ptc > 0 mapped to person and filtered to is_aca_ptc_eligible.
  • Positive PTC tax units/returns.
  • Assigned/used PTC amount.
  • APTC people per positive PTC tax unit, nationally and by state where support exists.
• Update the ACA take-up source-stage assignment/calibration to prefer tax units with more eligible people when needed, so the CMS APTC person target can increase without increasing SOI PTC return counts.
• Keep reported subsidized Marketplace coverage anchors preserved.
• Preserve deterministic seeded assignment.

Acceptance criteria

• Release diagnostics expose the CMS APTC people vs SOI PTC returns composition ratio.
• A focused source-runtime test covers a fixture where the solver must select fewer, larger tax units to satisfy both person-count and return-count constraints.
• The published-dataset check improves the CMS APTC-recipient person miss without worsening SOI PTC returns/amount beyond current tolerances.
• The implementation remains compatible with the current PE-US ACA take-up contract until Migrate ACA health-input gates after PE-US take-up contract changes #76 is unblocked.