Gate releases on input columns stuck at the engine default

[daphnehanse11]

Summary

Closes #257.

Adds a release gate that fails when any persisted PolicyEngine input column is constant at the engine's default value for every record. A default-valued column carries zero information — the H5 behaves identically without it — while looking populated to anyone inspecting the artifact. That is how the constant-40 weekly_hours_worked_before_lsr, the all-True takes_up_snap_if_eligible, and the all-CITIZEN ssn_card_type shipped in the current main-line release.

What the sweep found

Running the gate against the current main-line release (populace-us-2024-f0af251-703bd81a565c) flags 22 degenerate columns, including two previously untracked:

• s_corp_income constant zero (combined partnership/S-corp income carried in partnership_income in pre-PUF-support bases) — possibly relevant to Cross-check: Populace US 2024 net income & tax ~37–43% below Enhanced CPS (benefits match) #212
• spm_unit_tenure_type constant RENTER — misstates SNAP shelter deductions and SPM housing

Changes

• gates.py: default_valued_columns_gate — sweeps every column it is handed against per-column engine defaults. Constant-but-not-default columns pass and are reported (intentional broadcasts are modeling choices, not masked imputations). Bool columns never match numeric defaults (True == 1 must not excuse a stuck flag). Reviewed exclusions accept known offenders with a recorded reason: a stale exclusion (column now carries signal) fails the gate so the list cannot rot; a dormant one (column absent from this release line's surface) is only reported, since different release lines persist different column sets.
• PolicyEngine-US adapter: default_values(names) — engine defaults for non-formula input variables, enum defaults normalized to their stored member name; unknown and formula-owned names silently omitted so callers can pass a whole export surface.
• Release builder: _degenerate_input_signal_gate sweeps all entity tables (structural columns excluded), threaded into release gate failures, calibration diagnostics, and the build manifest alongside the health-input gate. Seeds US_DEGENERATE_INPUT_REVIEWED_EXCLUSIONS with the 22 known offenders, each naming its tracking issue (Carry hours-worked inputs through Populace US outputs #242/SNAP OBBBA work requirements: restore data inputs needed for household exposure analysis #248 hours, Carry SNAP reported amounts and take-up inputs through Populace US outputs #243 SNAP take-up, Dataset imputes 100% of the population as citizens with valid SSNs (breaks SSN/citizenship-conditioned policies, e.g. OBBBA CTC) #225 SSN/citizenship, Track SPM-specific input gaps: housing, WIC, school meals, child support, workers comp, and expense deductions #32 tenure/WIC, Use formula-constrained leaf imputation for deduction inputs #186 QBI flags, Track health input readiness for upcoming PolicyEngine-US model improvements #98 health take-up) — so existing releases keep building while any new degenerate column fails loudly, and the exclusion list is the visible debt register.

Verification

• Full populace-build + populace-frame suites pass (698 passed, 2 skipped), including 11 new gate tests, 2 new adapter tests, and the builder flow test.
• Gate run against the real cached main-line release H5: passes with exactly the 22 seeded exclusions in effect, zero stale, zero dormant.
• ruff check / ruff format clean on changed files.

Draft status

• Run a full fiscal refresh build end-to-end to confirm the gate result lands in build_manifest.json / calibration_diagnostics.json as expected on a real run.
• Decide whether the narrower fiscal-refresh base line surfaces additional offenders (its raw ASEC columns are unknown to the engine and skip automatically; its PE-facing surface is smaller).

🤖 Generated with Claude Code

[daphnehanse11]

@PavelMakarchuk context for review: verified against the current main-line release artifact (populace-us-2024-f0af251) — the sweep checks 178 input columns and catches 22 stuck at their engine default, all seeded as reviewed exclusions with tracking issues. Two of those were previously unknown: s_corp_income constant zero (possibly relevant to the #212 income shortfall) and spm_unit_tenure_type constant RENTER (SNAP shelter deductions).

One design choice worth your judgment: a stale exclusion (column carries signal now, e.g. SSN after #266 lands in a new base) fails the build until the exclusion is removed — strict by design so the list can't rot, but it turns an upstream fix into a red build for one cycle. Happy to soften to a warning if you'd rather.