Derive hours-worked inputs from ASEC reported hours

[daphnehanse11]

Summary

Closes #242; unblocks the SNAP work-requirements exposure channel tracked in #248.

The published dataset either omits weekly_hours_worked_before_lsr or ships it constant at the engine's 40-hour default, so PolicyEngine-US treats everyone as working 40 hours/week — SNAP's 30-hour general and 20-hour ABAWD tests pass universally and work-requirement reforms score ~zero through the hours channel.

This adds a hours_worked source stage following the #266 (immigration) template: manifest entry + shared runtime handler + frame transform + release gate + builder wiring.

What it does

Direct mappings of measured ASEC variables — identical to the retired enhanced-CPS pipeline, so parity is by construction, nothing imputed:

• weekly_hours_worked_before_lsr ← HRSWK (usual weekly hours last year; 0 for non-workers)
• hours_worked_last_week ← A_HRS1
• weeks_worked ← WKSWORK clipped to [0, 52]

Ownership handoff: the org_wages manifest entry no longer claims weekly_hours_worked_before_lsr — it never produced it (the retired pipeline sourced weekly hours from ASEC HRSWK directly, not from the ORG fit) — and consumes hours as predictors instead.

Healing behavior: the frame transform is idempotent when the hours surface carries signal, and recomputes from raw columns when weekly hours are constant — the published constant-40 landmine is repaired, not trusted.

Release gate (hours_worked_signal): weighted positive-hours share must land in [0.35, 0.62] (CPS ~50% of all persons worked last year) and mean weekly hours among workers in [30, 45] (BLS ~38.5). Threaded through release gate failures, calibration diagnostics, and both manifests, mirroring the immigration gate. The L0 refit export now requires all three columns nonconstant.

Verification

• 18 new tests (manifest declaration, derivation semantics incl. sentinel/NaN handling, frame integration incl. the constant-40 healing path, gate bands); full populace-build suite green.
• End-to-end against the current 572,780-person base: worked share 0.512, mean weekly hours among workers 37.3, gate green — real distributions, matching the retired enhanced-CPS surface (50.4% positive share).

Documented semantics caveat

SNAP's ABAWD standard is a monthly 80-hour test; HRSWK is an annual usual-hours measure. A person who worked half the year at 40 hrs/week carries 40 here. This matches how the retired pipeline fed the engine and is stated in the stage notes — week-by-week volatility is not observable in the CPS and is out of scope. The full ORG labor-market fit (tier 2) is tracked in #291.

Merge-order note

After this merges, the weekly_hours_worked_before_lsr reviewed exclusion in #286's degenerate-input gate becomes stale and should be removed (the gate will then permanently enforce non-degenerate hours). If #286 merges first, that's a one-line follow-up here; if this merges first, I'll drop the exclusion in #286.

🤖 Generated with Claude Code

[daphnehanse11]

Rebased onto current main (8b1795a, was 14 behind) and squashed the two commits — the indentation-restore commit was fighting main's reformat of source_stages.json and made every rebase conflict on the whole file.

Beyond conflict resolution, one intended change: dropped the now-stale weekly_hours_worked_before_lsr reviewed exclusion from US_DEGENERATE_INPUT_REVIEWED_EXCLUSIONS — the merge-order follow-up flagged in the PR description (#286 merged first, and this PR makes the column non-degenerate, so leaving the exclusion would itself fail the cannot-rot gate).

Conflict resolutions were all additive interleaves with main's newly merged take-up (#315) and eCPS-parity (#316) wiring: both gate sets now run in sequence in main(), and hours_worked_gate threads through _release_gate_failures/_build_manifests alongside degenerate_input_gate/ecps_parity_gate. Verified the rebased net diff against the original with interdiff — no other semantic drift.

Local verification: full populace-build suite green (uv run --package populace-build --extra us pytest), ruff check + format clean on touched files.

🤖 Generated with Claude Code