Split unsplit dividend totals by SOI qualified share

[MaxGhenis]

Summary

Fix the qualified / non-qualified dividend split inversion in normalize_dividend_columns.

When a record carries only a dividend total (e.g. CPS DIV_VAL, which reports no qualified/non-qualified breakdown) and no observed component split, the function dumped 100% of the total into non_qualified_dividend_income and left qualified_dividend_income at $0. Because this runs on the merged persons (pipelines/us.py:10854,11090) where every CPS-native dividend row has a total but no split, it zeroed qualified dividends on the entire CPS spine and inverted the national split.

Measured on the collapse-scaffold RC (audit dashboard, scored via microdf weighted .sum() against the SOI targets):

	qualified	non-qualified	qualified share
collapse RC (before)	$102B	$465B	18% (inverted)
eCPS	$352B	$148B	70%
PUF 2015 source (E00650/E00600)	$204B	$57B	78% (correct)
SOI target (qualified_dividend_income)	$291B	—	—

qualified_dividend_income was −65% vs its $291B target; the total dividend was roughly right, so the error was purely in the split.

Fix

Allocate an unsplit dividend total by the SOI qualified share (UNSPLIT_DIVIDEND_QUALIFIED_SHARE = 0.78, basis: SOI 2015 PUF E00650/E00600 = $204.0B/$260.9B = 0.782) instead of defaulting the whole amount to non-qualified. Applies in the two branches that previously produced all-non-qualified output (has_qualified and has_non_qualified total-only rows, and the no-components else branch).

This mirrors how the incumbent eCPS splits CPS DIV_VAL via a qualified fraction. The 0.78 flat share is a documented first-order constant; a future refinement is an income-conditional share.

Blast radius

• normalize_dividend_columns has 4 call sites. PUF data is unaffected — it carries the E00650/E00600 split, so the total-only branch never fires there. Only merged-persons rows that have a dividend total but no observed split (the CPS-native spine) change.

Tests

• Updated test_normalize_dividend_columns_coalesces_sparse_total_aliases_by_row (row 0 now splits the unsplit total 78/22 instead of 0/100; rows with observed components unchanged).
• Added test_normalize_dividend_columns_splits_unsplit_total_by_qualified_share.
• test_normalize_dividend_columns_prefers_atomic_components_over_totals still passes (observed components are preserved, never reallocated).

Validation note

Unit-tested. Full national validation requires a data rebuild (the qualified/non-qualified split feeds calibration + tax computation); expected to move qualified_dividend_income from ~$102B toward the ~$291B SOI target and deflate non_qualified_dividend_income correspondingly. Part of the calibration-surface fixes tracked in #200 (this is the confirmed, self-contained one; the interest seam, wage/NIPA target profile, and tax-expenditure target handling remain there).