Fix nonsense relative error on zero-target rows

[PavelMakarchuk]

Problem

The Target diagnostics table rendered astronomical percentages like 5,191,840,415,890% for SOI cells whose target is genuinely $0 (the under-$1 AGI band, several EITC-by-AGI cells). Relative error against a $0 target is undefined; the code fell back to the raw dollar miss and displayed it as a percent. These bogus rows also dominated the worst-fit sort and made the calibration look broken.

Fix

• relativeError() returns null when the target is 0.
• enrichTargetRow nulls relative_error/abs_relative_error for zero-target rows (overriding any value the producer published), exposes the dollar miss as abs_error, and sets target_is_zero.
• The targets table shows " vs 0" for those rows instead of a percent, and they sort to the bottom of the relative-error ranking.

Verified live against the current release: zero-target rows no longer show a percent; the remaining large percentages all have non-zero targets (e.g. a $535K target vs a ~$10B estimate) — real calibration misses, left untouched.

Tests

bun test — 9 pass (added a zero-target regression). tsc --noEmit clean.

🤖 Generated with Claude Code

[PavelMakarchuk]

Second commit on this branch also canonicalizes the dotted ledger target names so both naming schemes are supported.

This release names targets with dot paths (irs_soi.ty2022.table_2_5.eitc_by_agi_children.no_qualifying_children.25k_to_30k.eitc_total) instead of the slash form (nation/irs/…). The parser only understood slashes, so the variable browser showed 5,229 ungrouped rows with empty geography/breakdowns.

parseTarget now dispatches by name shape and parseDottedTarget decomposes the dotted scheme into the same {source, variable, measure, geography, breakdown}:

• 5,229 names → 83 clean variables across all 11 families (IRS SOI, JCT, CBO, Census STC, CMS ×3, HHS, SSA, USDA SNAP), 0 unparsed.
• …_amount/…_total → amount, …_returns/…_claims → count (so a variable's amount and count rows group separately, e.g. IRS SOI / eitc · amount vs · count).
• AGI bands humanized (under_1 → "AGI <$1", 25k_to_30k → "AGI $25k–$30k"), children (no_qualifying_children → "0 children"), geography populated — income-band and children facets work.
• Pure-measure leaves take the descriptive token (jct.…salt_deduction.revenue_loss → "salt deduction", census_stc.…individual_income_tax_collections.al.t40.collections → "individual income tax collections", skipping the t40 code).

Slash names unchanged. 4 new tests (13 total).

[PavelMakarchuk]

Third commit: variable-level comparison across naming schemes.

Comparing the 457-target slash release (A) to the 5,229-target dotted release (B) showed "0 in common" — the per-target diff matches on raw names, which differ entirely between the two conventions. buildComparison now also emits a variable_comparison that groups each release by a scheme-independent canonical variable key and reports per-variable fit (n, within-10%, mean |err|) for variables in both releases.

Across the migration boundary this recovers 11 shared variables (the 5 JCT tax expenditures + 6 IRS income lines), e.g.:

Variable	A within 10%	B within 10%	Δ
adjusted gross income	89% (70)	60% (573)	−29%
taxable pension income	100% (2)	18% (124)	−82%
SALT deduction	100% (1)	100% (1)	0

The compare view shows a "Variable-level fit (A → B)" table always, and the per-target improved/regressed tables only when common targets exist.