Baseline as a first-class attribute of every score

Max (8/2): "we should have the baseline as a first class attribute of the score in our db." The day supplied four independent proofs: PWBM tables scored vs TCJA-extension; JCX-29-25 scored vs current policy (twin of JCX-35-25 vs present law); TPC's CTC set scored vs current law + Senate Title VII; and PE's own two-child-limit run, where the 2026 current-law baseline already contains AB2025's abolition so "removal" had to be measured against a reinstated counterfactual. A score's meaning IS its (reform, baseline) pair — baseline can't live half-inside ReformRef.baseline JSON and half in conditions.baseline_policy strings.

Schema change

1. baselines registry table: baseline_key (hash) · label (closed slug vocabulary, extend deliberately: current_law | current_policy | tcja_extension | pre_ab2025 | current_law_plus_senate_title_vii | …) · description · framework · spec_json (nullable — many external baselines are named conventions, not expressible parameter dicts) · provenance (the source's own words defining it).
2. baseline_key column on external_scores, pe_results, and pe_exhibits (FK to registry). Claim-id hashing unchanged (ReformRef keeps carrying baseline content for the hash) — this is an additive queryable projection, no id migration.
3. comparisons view exposes both the claim's baseline label and the PE result's actually-run baseline. Guard: rows where claim-baseline ≠ pe-baseline can never render as plain agreement — auto-annotate constructed (or concept_mismatch) unless a registry equivalence says otherwise. App (#9): baseline chip on every row, same pattern as vintage chips.
4. Migration of existing data: Urban rows → current_law (2023); fullpart/solo exhibits → baseline current_law with input-override reform; harvest ingest maps the staged baseline_policy / baseline_hint / conditions.baseline fields into registry entries (PWBM, JCT, TPC ones enumerated in COLLATION.md).
5. PE-side rule: every run records the baseline it actually executed (e.g. tonight's two-child rows: reform=pre_ab2025→current_law direction made explicit rather than implied by sign).

Grounding artifacts: ~/populace-sotsn-takeup/results/uk/two_child_reinstate_2026.json (the reinstated-counterfactual construction), COLLATION.md §ingest-worklist item 3 (baseline variants are load-bearing).

Read COLLATION.md §ingest-worklist item 3 and the two_child_reinstate_2026.json grounding artifact first, then trace the external_scores, pe_results, pe_exhibits, and comparisons data paths. Done means baselines are registry-backed, existing data is migrated, executed and claimed baselines are exposed and mismatches are annotated, and every score displays its baseline.