_drop_formula_variables drops employment_income, zeroing it in CI builds

[baogorek]

Bug

_drop_formula_variables() in ExtendedCPS (introduced in PR #554, commit 7025f6a) drops employment_income and self_employment_income because they have adds. But the raw CPS H5 only stores employment_income — not employment_income_before_lsr. So income data is silently lost during the ExtendedCPS build.

Chain of failure (CI builds from scratch)

1. CPS_2024 H5 stores employment_income
2. ExtendedCPS_2024.generate() calls _drop_formula_variables() → drops employment_income (has adds)
3. ExtendedCPS H5 now has no employment income variable
4. EnhancedCPS_2024 loads ExtendedCPS → no employment income
5. Microsimulation computes employment_income via adds → needs employment_income_before_lsr → not in H5 → defaults to 0

Same issue for self_employment_income / self_employment_income_before_lsr.

Evidence

# CPS H5 contents
employment_income: shape=(50692,), sum=1.84e+09        # ← present
employment_income_before_lsr: NOT IN H5                # ← missing

# _drop_formula_variables behavior
employment_income:            would_be_dropped=True     # has adds
employment_income_before_lsr: would_be_dropped=False    # input variable

Why it's not caught locally

Local EnhancedCPS data downloaded from HuggingFace was built before _drop_formula_variables existed, so it still has employment_income_before_lsr. Only fresh CI builds (which regenerate from scratch) hit this.

Discovered via

PR #570's new sanity tests caught employment_income summing to 0 in CI.

Possible fixes

• Don't drop variables with only adds/subtracts (only drop those with actual formulas)
• Or rename employment_income → employment_income_before_lsr before the drop step