Implement ACA marketplace take-up input materialization

[daphnehanse11]

Context:
Issue #70 identified that the published Populace US 2024 H5 has degenerate ACA take-up input signal: takes_up_aca_if_eligible is true for every tax unit. PR #73 declares the missing aca_marketplace_inputs source stage, but does not implement the constructor.

Goal:
Materialize ACA marketplace take-up in Populace instead of inheriting a constant/defaulted column from the base H5.

Suggested implementation:

• Use CPS reported Marketplace/subsidized Marketplace coverage as true anchors at the tax-unit level.
• Assign simulated take-up to non-reported tax units using stable seeded tax-unit draws and the ACA take-up rate input.
• Calibrate the assignment to CMS Marketplace/APTC enrollment targets, preferably by state where support allows.
• Preserve reported true anchors during calibration.
• Keep the implementation independent of policyengine-us-data.

Acceptance criteria:

• The resulting H5 has nonconstant ACA take-up signal.
• A deterministic test covers reported anchors and stable seeded assignment.
• A calibration/diagnostic test covers matching a small synthetic CMS target.
• The US fiscal refresh health-input gate passes on a constructed fixture with realistic take-up variation.

Related:

• Umbrella: Investigate ACA take-up and plan-choice inputs driving high PTC estimates #70
• Source-stage declaration: Declare ACA marketplace input source stage #73
• Healthcare target mapping: Fix healthcare fiscal target variable mappings #68
• Degenerate-input guard: Guard US fiscal refresh against degenerate ACA inputs #71

[MaxGhenis]

PR #83 landed the first runtime implementation for this source stage on main (be8d56a): tax-unit aggregation from CPS person Marketplace coverage, seeded take-up assignment, reported-anchor preservation, weighted calibration to CMS count targets, and loud failures for missing eligibility/weight columns. The full H5 acceptance criterion is still pending until the source-runtime output is wired into the production source-to-artifact build and a fresh dataset is generated.

[daphnehanse11]

Validation note from current main (843f1e0 locally):

• The current active ACA facts compile into 102 cms_aca targets: 51 marketplace enrollment targets and 51 APTC-recipient targets.
• _aca_source_target_tables(...) currently feeds the ACA source runtime only the APTC-recipient table (cms_aca_aptc_recipients_by_state), whose 2024 national target sums to 19,743,689.
• Running the ACA source stage against the current downloaded base H5 gets through the builder's support_clip stop point, but the base frame only has 4,489,655 weighted eligible/APTC-support people. The state take-up rates therefore clip to 1.0, and every eligible support record takes up ACA, but the frame still cannot reach the target under the current underweighted base.

This looks consistent with the broader base-weight/mass blocker I noted on #67, not a new ACA formula bug. I would keep this open until a fresh release build can be run from a national-scale base H5 or non-conserved mass calibration and the APTC/enrollment targets can be checked in the emitted diagnostics.