Fail the eCPS parity gate on stale exemptions; remove six rotted entries

[daphnehanse11]

Summary

parity_gate skipped names in the known-gaps exemption register (ecps_parity_known_gaps.json) unconditionally, with no staleness check — unlike its two sibling gates, which both fail on stale reviewed exclusions (default_valued_columns_gate at gates.py:707, source_stage_input_coverage_gate at gates.py:1499). That let register entries rot silently once the candidate started producing the exempted layer, which is exactly what happened to takes_up_medicaid_if_eligible when PR #334 started materializing the column.

Gate change

• parity_gate now fails on any stale exemption — a known_gaps name whose candidate nonzero share is positive — with a message to remove or re-reason it.
• Exemptions for layers the reference never populates are classified dormant and only reported (different reference vintages populate different layers), matching the sibling gates' treatment.
• Details gain stale_exemptions and dormant_exemptions, flowing into the release manifest through _ecps_parity_gate unchanged.

Register audit: six entries removed

Checked every exemption against the current release pipeline (rebased onto main after #293/#294/#345/#330/#344/#348/#350/#358 landed). The parity gate runs on the base frame after the immigration, take-up, hours-worked, SNAP take-up, and eligibility-input transforms, and each transform's own signal gate hard-fails unless its columns are present and non-constant — so six exempted layers are deterministically populated at parity-gate time (all confirmed in PolicyEngineUSEngine().variables(), so they are measured):

Entry	Produced by
takes_up_tanf_if_eligible	contract-seeded take-up (#315)
takes_up_eitc	contract-seeded take-up (#315)
ssn_card_type	immigration stage (#266)
immigration_status_str	immigration stage (#266)
hours_worked_last_week	hours-worked stage (#293, merged without removing its register lines)
weekly_hours_worked_before_lsr	hours-worked stage (#293, merged without removing its register lines)

The register was seeded (#316) from a parity run of an artifact predating the first four; #293 merged after the seeding without cleaning up its two. hourly_wage stays: it is an org_wages output and that stage is not executed anywhere yet. #294 (SNAP) and #350 (eligibility inputs) removed their own register lines at merge, so nothing further rotted. The remaining entries are safe: the other source stages declaring exempted outputs (SCF wealth, ORG wages, SIPP tips, …) are manifest declarations no tool currently executes.

Static CI guard

test_register_exempts_no_release_produced_layer pins the statically-knowable release-produced set so a stale entry fails in CI rather than first on a release build:

• immigration outputs, hours-worked outputs, SNAP take-up flag, and eligibility-input outputs (the deterministic stage-output constants — Derive hours-worked inputs from ASEC reported hours #293 slipped through exactly because the guard didn't exist yet), and
• take-up flags whose contract treatment is in {seed, count_calibrated} (per Max's Assign SNAP take-up from reported receipt and the FNS participation rate #294 review: each producing PR removes its own register line; keyed on the treatment string rather than Materialize Medicaid take-up via anchored count-calibration to CMS state snapshots #334's count_calibrated_take_up_programs() helper so this branch needs nothing that isn't on main yet, and picks up the Medicaid flip automatically in either merge order).

Interaction with #334

#334 deletes the takes_up_medicaid_if_eligible exemption from the same register when it materializes Medicaid take-up. The hunks are disjoint, so the two PRs merge cleanly in either order, and they compose consistently: on current main the candidate does not yet produce that column (treatment rate_unsourced), so the entry stays; once #334 lands, its own deletion is exactly what the new enforcement (and the static guard, via the treatment flip) demands. The medicaid case is the motivating example for this change. Closes the follow-up from the #334 review.

Tests

• Five new TestParityGate cases: stale fails and names the variable; stale fails even when the reference lacks the layer; dormant passes but is reported; a stale exemption does not mask real gaps; clean-exemption details stay empty.
• The static register guard above in test_us_parity_reference.py.
• test_gates.py, test_us_parity_reference.py, and test_us_nonzero_shares.py pass (141 tests); ruff check/format clean.

Note: the runtime enforcement means the next release build fails if any exemption not statically verifiable here turns out to be populated on the real base H5 — that failure names exactly which entry to remove or re-reason, which is the intended behavior.

🤖 Generated with Claude Code

[daphnehanse11]

Follow-up from Max's #294 review: the static register guard now derives its produced set from contract treatments ({seed, count_calibrated}) rather than seeded_take_up_programs(), so programs materialized by count calibration (Medicaid in #334, CHIP next per the contract followup) are covered in CI instead of first failing on a release build.

Deliberately keyed on the treatment string rather than #334's count_calibrated_take_up_programs() helper, which doesn't exist on main yet — this branch stays green now, and the guard picks up the Medicaid flip automatically in either merge order (#334 deletes its register line in the same commit that flips the treatment). A future PR that flips a treatment but forgets the register line fails this test.

[MaxGhenis]

Approve

Verified the removals and the fail-on-stale mechanism against origin/main.

Every removed exemption is genuinely rotted. I reconstructed the static-test produced set on origin/main (immigration + hours + eligibility + SNAP take-up + take-up-contract programs treated seed/count_calibrated) and intersected it with the register. The stale set is exactly the six this PR removes — no more, no less:

removed	why it is produced now
hours_worked_last_week, weekly_hours_worked_before_lsr	in US_HOURS_WORKED_OUTPUT_COLUMNS (the #293/#242 hours stage)
immigration_status_str, ssn_card_type	in US_IMMIGRATION_OUTPUT_COLUMNS
takes_up_eitc, takes_up_tanf_if_eligible	populace_treatment: "seed" in take_up_contract.json — the only two produced take-up flags

Because the main loop only fails when a reference-populated layer is candidate-all-zero, removing an exemption for a layer the candidate populates (cand_share > 0) produces no failure — the gate passes without it. Keeping it now fails. Register shrinks 103 → 97.

Kept entries are correctly kept. The remaining take-up exemptions (takes_up_medicaid_if_eligible, medicare, ssi, dc_ptc, head_start, early_head_start, housing_assistance) all carry rate_unsourced/near_universal, i.e. not produced, so they're legitimately still exempted. takes_up_medicaid_if_eligible is rate_unsourced on main — that's why it's the docstring's illustrative rot case rather than a removal here.

Fail-on-stale matches the #286 "cannot rot" pattern — literally. The new stale/dormant block is a direct port of the reviewed-exclusions structure already sitting above parity_gate in gates.py (the #286 degenerate-input gate: "Stale reviewed exclusions — the leaf is produced now, remove the exclusion"). The parity adaptation is correct on two axes: stale keys off candidate share (a produced layer makes the exemption a lie, even when the reference lacks it), dormant keys off reference share and is guarded by name not in stale. gap_count is snapshotted before the stale append, so stale exemptions don't inflate the gaps metric — test_stale_exemption_does_not_mask_real_gaps pins exactly that.

SNAP-train interaction is clean. No column carried by the recently-merged SNAP train (#350/#352/#353) — is_disabled, is_blind, is_full_time_college_student, own_children_in_household, veterans_benefits, nor any snap/abawd/pregnan input — is left in the register as a stale exemption; those producing PRs removed their own lines. takes_up_snap_if_eligible (out_of_scope in the contract, produced by the #294 stage) is correctly covered by the unconditional US_SNAP_TAKE_UP_OUTPUT_COLUMN term rather than by treatment. Post-removal the static test yields an empty stale set.

One cross-PR dependency worth calling out: #334 flips takes_up_medicaid_if_eligible to count_calibrated. Once both land, whichever merges second must ensure the medicaid register line is gone, or test_register_exempts_no_release_produced_layer fails — which is exactly the forcing function this PR installs (the comment already anticipates it). Confirm #334 drops that register entry.

The trailing test_missing_issue_is_refused hunk is a cosmetic one-line reformat. CI green (test + wheels, 3.13/3.14), MERGEABLE/CLEAN. Not merging (review-only per request).