PUF donor home_mortgage_interest carries the E19200 total-interest concept (+7.6% vs the Table 2.1 mortgage-only leg)

[MaxGhenis]

Found root-causing #511 (mortgage dollars +29.5% on certified O-1). Second of the two concept gaps behind that row; the first (target-side gross-vs-deducted) has its own PR.

The defect

The donor column home_mortgage_interest in the pinned PUF artifact (irs-soi-puf/1.8.0/puf_2024.h5, consumed at base-o per its summary puf_h5/puf_sha256=7669f5b5…) carries the E19200 lineage — Schedule A "Interest paid deduction, Total" — not the home-mortgage-only concept the SOI Table 2.1 target measures. The repo's own evidence notes document the construction (release_input_coverage_manifest.json investment_interest_expense exclusion; mortgage_conversion stage notes): archived puf.py L654 maps E19200 → interest_deduction, L1548-1550 assigns the identical amount as deductible mortgage interest, and utils/mortgage_interest.py L268-320 derives investment_interest_expense only as the (zero) residual. The artifact has 499,045 all-zero investment_interest_expense values.

At TY2015 (SOI Pub 1304 Table 2.1, all itemized returns):

line	amount
Interest paid deduction, total (=E19200)	$304.461B
Home mortgage interest (FI + individuals) — the target concept	$283.004B
Deductible points	$1.274B
Qualified mortgage insurance premiums	$6.287B
Investment interest expense	$13.895B

So every imputed home_mortgage_interest draw is concept-inflated by ×1.0758 (global; the ratio varies by AGI band — investment interest concentrates at the top).

Receipts on certified O-1 (buildo-sparse-rmloss100-22bd902)

Itemizer-masked masses at final weights vs the $186.31B aged Table 2.1 fact:

step	mass	rel err
gross home_mortgage_interest (the #511 row as shipped)	$241.27B	+29.5%
engine-capped deductible_mortgage_interest_tax_unit (target-side PR)	$214.20B	+15.0%
÷1.0758 donor concept scale (this issue)	$199.11B	+6.9%

Residual +6.9% (count +2.45%, mean +4.3%) is within normal solve slack — carriers are broad (5,762; top-1 2.6% per diagnose_us_target_support), so no #481-style tail story on this row. Donor realized ETL uprate on the column ex-aggregate-rows: ×1.374 ($304.46B → $418.25B at 2024, measured via puf_tax_unit_donor_from_arrays on the pinned artifact).

Fix shape (#487's variable-specific donor mechanism, applied as a concept carve)

Root fix in the irs-soi-puf ETL: carve E19200 into a mortgage leg and an investment-interest(+points+QMIP) leg using SOI TY2015 Table 2.1 shares — ideally by AGI band — emitting home_mortgage_interest at the A19300-family concept and a real investment_interest_expense column (reversing the SOURCE-UNAVAILABILITY exclusion with the SOI table as the evidence base). Interim populace-side option: scale the donor column (and the first/second_home_mortgage_interest structural legs, which carry the same total) by 283.004/304.461 = 0.9295 at donor load, factor ledgered — the #486 repair precedent.

Interaction note: the JCT deductible_mortgage_interest reform neutralizes interest_deduction; while investment_interest_expense is all-zero that row measures the mortgage-only concept by accident. If the carve un-zeroes investment interest, the reform should neutralize deductible_mortgage_interest instead — otherwise the JCT row re-inflates by the carved leg.

Relates #511, #487 (mechanism class), #486 (repair precedent). Full forensics on #511.

[MaxGhenis]

Claiming (same lane that closed #511/#517, which is now merged). Implementing the interim populace-side carve: scale the donor's E19200-lineage columns (person home_mortgage_interest + first/second_home_mortgage_interest structural legs) by the SOI TY2015 Table 2.1 mortgage-only/total-interest ratio (283,004,465k / 304,461,163k ≈ 0.9295) at the puf_tax_detail donor seam, factor defined from the two ledgered SOI numbers with citation. Balances/origination untouched; investment_interest_expense stays zero (root ETL carve remains open here). Implementation dispatched to sol; review-cycled PR to follow. The #517 register ratchet (0.20 → 0.15) activates once this lands and a run holds.

[MaxGhenis]

Interim carve is merged (#525, squash): the puf_tax_detail donor now scales its E19200-lineage columns by 283,004,465/304,461,163 at donor construction, with the checkpoint-schema bump (v2) so pre-carve QRF checkpoints cannot fit/draw under carved code, and the loader now validates target-checkpoint schema. Two sol review rounds absorbed (checkpoint invalidation, fail-open regression, target-schema validation). Next certified build should land the Table 2.1 mortgage amount row in the ~+7% band; the #517 register ratchet (0.20 → 0.15) activates once a run holds. This issue stays open for the root ETL carve (per-record E19200 split by AGI band, un-zeroing investment_interest_expense, and switching the JCT reform to neutralize deductible_mortgage_interest).