Materialize Marketplace selected-plan benchmark ratios and metal-level diagnostics

[daphnehanse11]

Context:
Issue #70 found that selected_marketplace_plan_benchmark_ratio is 1.0 for every tax unit in the published Populace US 2024 H5. PR #73 declares selected-plan ratio construction in the aca_marketplace_inputs stage, but does not implement it.

Goal:
Generate Marketplace selected-plan benchmark ratios with real signal, then validate them against CMS metal-level Marketplace data.

Suggested implementation:

• For Marketplace take-up units, compute selected_marketplace_plan_benchmark_ratio from reported non-Medicare private premiums, assigned ACA PTC, and SLCSP.
• Clip or flag outliers using a reviewed support range, such as 0.5 to 1.5.
• Keep non-Marketplace units at the neutral default when the variable is not consumed.
• Add diagnostics comparing below-benchmark or bronze-like selections to CMS metal-level APTC enrollment targets where available.
• Record outlier counts and distribution summaries in the build diagnostics or release manifest.

Acceptance criteria:

• Marketplace takers have nonconstant selected-plan ratios.
• Unit tests cover ratio construction, clipping, zero/invalid SLCSP handling, and neutral defaults for non-takers.
• Diagnostics report state/national plan-choice shares against CMS metal-level targets.
• The health-input gate catches all-1.0 ratios before release.

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
• Take-up materialization sibling: Implement ACA marketplace take-up input materialization #74
• Degenerate-input guard: Guard US fiscal refresh against degenerate ACA inputs #71

[MaxGhenis]

PR #83 landed the selected-plan ratio runtime prefix on main (be8d56a): compute_ratio over declared numerator/denominator columns, neutral defaults off-domain or invalid SLCSP, support clipping to the reviewed 0.5-1.5 interval, and expression calibration support for below-benchmark plan-choice targets. Full release diagnostics against CMS metal-level targets still need to be emitted by the source-to-artifact build before this issue can close.

[daphnehanse11]

Validation note from current main (843f1e0 locally):

The generated active 2024 ACA facts provide APTC-recipient and marketplace-enrollment targets, but they do not provide cms_aca_bronze_aptc_consumers_by_state / bronze-or-metal selected-plan target tables. Because of that, the release builder follows its fallback path:

stop_after = "support_clip"

and does not run the selected-plan benchmark-ratio calibration step that requires the bronze target table.

So #75 still looks open after #83/#85/#86: we have the runtime machinery, but the active release facts/diagnostics still do not exercise the bronze/metal selected-plan calibration or emit the full CMS metal-level acceptance diagnostics.