Implement in-tree runtime handlers for declared US source-stage operations

[daphnehanse11]

Problem

After the spec-only country packages change (da8e14a), US source stages are declared in source_stages.json but most cannot execute: us_source_operation_handlers() covers only 6 operation kinds (all ACA-related plus support_clip/disaggregate_aggregate_records). Every other stage's core operations have no handler anywhere in the repo:

• org_wages: fit_labor_market_models — blocks hours-worked inputs (Carry hours-worked inputs through Populace US outputs #242, SNAP OBBBA work requirements: restore data inputs needed for household exposure analysis #248)
• scf_wealth: replace_sentinels, derive, fit_weighted_qrf, head_carry
• sipp_tips: fit_tip_income_model, zero_when_false
• meps_esi_premiums: assign_by_plan_type
• prior_year_income: join, replace_sentinels
• mortgage_conversion: all three operations
• acs_rent: aggregate_person_to_household, fit_weighted_imputer, zero_when_false, head_carry
• vehicle_assets: fit_vehicle_model, fold_into

Consequence: the published base H5s cannot be reproduced from this repo — only recalibrated. Every input-parity issue (#242, #243, #244, #249) implicitly depends on being able to re-run source stages.

Proposal

Implement shared runtime handlers for the declared operation kinds, prioritized by downstream need:

1. fit_labor_market_models (org_wages) — hours inputs are the SNAP work-requirements blocker
2. fit_weighted_qrf + head_carry + replace_sentinels/derive (scf_wealth, acs_rent share these)
3. The remaining stage-specific fits

The immigration/SSN source stage (#266) established the pattern: manifest entry + shared runtime handler + tests. Handlers should be generic interpreters of manifest parameters (like the existing ACA handlers), not country-specific Python.

Context

Found while auditing SNAP work-requirements readiness: the manifest declares weekly_hours_worked_before_lsr as an org_wages output with full provenance, but nothing in-tree can produce it.