Add the work_experience_inputs stage: industry recodes and worked-last-year on every US person record

[MaxGhenis]

Fixes #719.

What

A new post-clone source stage, work_experience_inputs, puts three measured CPS work-experience attributes on every US person record:

column	source	semantics
detailed_industry_recode	ASEC WEIND	industry of longest job by detailed groups (0:23; 22 = Military, 23 = Never worked; 0 = NIU)
major_industry_recode	ASEC WEMIND	industry of longest job by major industry groups (0:15)
worked_last_year	ASEC WKSWORK > 0	the official universe condition of the work-experience longest-job recode block

These are the reference-period-coherent siblings of detailed_occupation_recode (POCCU2, occupation of longest job): the survey-week A_DTIND/A_MJIND recodes would describe a different job than the occupation the model already carries. Names and universes are verified against the official ASEC public-use data dictionary (asec2024_ddl_pub_full.pdf, person entries at positions 326/329).

The frozen census_cps person inputs never carried WEIND/WEMIND, so a new sidecar (work_experience_source.py) restores them for every pooled income year from the three SHA-pinned official ASEC archives via exact PERIDNUM joins — the same archives and pattern as the education-assistance/PAW_TYP sidecars, with the stage's own measured audit pins. Nothing is imputed; both support clones of a source person inherit identical values, as the occupation recode does. The outputs also join the ACS-spine transfer surface automatically (one bounded group, three targets), so ACS-spine persons receive QRF-transferred values like every other pool operator output.

Why WKSWORK > 0 rather than WORKYN == 1

Measured on the pinned archives: WEIND holds a worker code (1–22) iff WKSWORK > 0 with zero violations in all three years (worked rows 73,186 / 73,471 / 72,460), while WORKYN = 1 under-covers that universe by 642 / 572 / 594 allocation rows per year that carry positive weeks and a real industry (WORKYN = 1 never appears without positive WKSWORK). The loader enforces these identities at load time and the derive handler re-enforces them on the pool surface with zero tolerance.

Wiring

• Manifest: source_stages.json + authoritative spec/sources.yaml (+ sources.schema.json kind), US_DONORS, US_STAGE_NAMES, allowed operation kinds, central handler registry.
• Pool: operator contract/order/lambda, late-producer inventory, operator_boundary output family, stacked-spine battery metrics (categorical_tvd recodes, boolean_incidence indicator) and stage/spec/sidecar bindings; pinned partition counts move 19→20 groups, 70→73 late targets, 118→121 transfer / 131→134 terminal surface.
• Raw checkpoint: asec_raw_stage schema version 4 restores WEIND/WEMIND alongside ED_VAL/LKWEEKS/PAW_TYP (industry-blind artifacts fail loudly); build_us_puf_support_base.py loads the sidecar from the shared --asec-education-source archives and runs the stage with its gate on both builder paths.
• Release: POST_REFERENCE_ECPS_REQUIRED_INPUTS + regenerated coverage manifest (163 → 166 required); L0-refit export requirement.
• Tests: sidecar (16), stage (24), plus fixture/enumeration updates across the checkpoint, plan, multispine, and country-spec suites.

Dependency — why this is a draft

The engine-side input variables land in PolicyEngine/policyengine-us#9311 (31/31 checks green, mergeable). Microcosm checks release columns and transfer targets against the locked PolicyEngine-US (uv.lock pins 1.764.6, unchanged since July), and tools/generate_us_bundle_from_constants.py projects us/spec/*.yaml + engine_abi.lock.json from engine metadata, so with the current lock exactly two categories stay red:

1. Engine-leaf checks: test_us_multispine_pool.py::test_every_pool_transfer_target_is_an_installed_engine_input_leaf, test_release_input_coverage.py::TestShippedManifest::test_manifest_is_current, and the whole of test_us_pool_input_consumers.py (validating the three new CLIENT_SURFACE allowlist entries looks the variables up in the locked engine).
2. Bundle projection + attestation (one regeneration pass): test_us_spec_bundle.py, test_imputation_lineage_spec.py, test_us_bundle_core_contracts.py (catalog entries), test_spec_engine_* (battery/imputation/inventory/field-usage/identity/loader/legacy-adapter/stacked-authority projections and the reviewed digest vector in inventory_coverage.py, including source_stages == 37), and test_us_multispine_pool_tool.py (the constants-vs-bundle battery contract and the stacked goldens that fold bundle identities). The generator cannot run until the installed engine knows the variables.

Everything else passes locally: the new-stage suites, checkpoint/builder/L0-export fixtures, the late-DAG and stacked-authority pins (20 groups / 73 late targets / 134-target terminal surface), spine-blindness classification (runtime graph 65 → 67), the CLIENT_SURFACE consumer review for the three columns, and the BE/UK spec digests (re-pinned: the seed-protocol attestation digests the source of us_runtime/source_runtime.py, where the derive handler registers). test_us_trade_imdb_bulk.py::test_build_cli_success_replaces_prior_publication_completely failed once in a parallel full run and passes in isolation — unrelated.

Moving the pin to ≥1.817 is itself a procedure — tracked as #746 (generated-variable audit re-pin, take-up-contract totality for four new engine flags, and pe-us's rename of would_claim_wic → takes_up_wic_if_eligible, which migrates the existing wic_claim stage). Sequence: merge #9311 → #746 → rebase this branch, run the generator, re-cut the attestation vector and goldens, un-draft.

No frozen census_cps artifacts, HF destinations, or release publication are touched.

🤖 Generated with Claude Code