Fold three concept-spelling duplicate lineages; defer entity-drift splits

[MaxGhenis]

Curation pass over the 10 duplicate-lineage families flagged after the v3 catalog rebuild (follow-up chip from #128's merge disposition).

What folds

Three same-dimension concept-spelling pairs — identical geography and entity key, so alias inheritance is in-dimension and no observation moves across identity slices:

Retired spelling (noted retire event)	Surviving lineage
dol.eta.initial_claims.sa (330ec66b)	830e68bb
us.census.housing_starts.total_saar (d542cb8d)	d0212a1e
us.frb.industrial_production.total.mom_sa (b5a74252)	d5634fc4

Catalog: 209 → 206 series. Registry: +3 noted retire appends (216 entries), append-only. No surviving UUID changes; 171/171 observations preserved; all 83 docket-pinned UUIDs remain live. The identity-map sha256 anchor moves in this same diff, per its contract.

What does not fold, and why

The seven entity-drift families (May/June observations disagreeing on entity) would require moving observations across entity keys — exactly what dimension-scoped alias inheritance forbids, and a disposable-copy experiment confirmed regeneration restores the split. They are recorded with per-family blockers in docs/catalog-curation-backlog.md, alongside the us.dol.initial_claims.sa weekly/monthly split that awaits upstream cadence-metadata correction (the "monthly" rows carry week-ending identifiers and weekly-thousands values).

Verification

• pytest tests/test_build_series_catalog.py: 153 passed (on top of Builder plans re-enrichment; succeeds schema errors are clean findings #145's builder)
• Regeneration byte-idempotent: rebuild reproduces committed catalog + registry exactly (minted=0, superseded=0, suspects=0, ambiguous_aliases=0)
• build_series_catalog.py --check: current at 206
• ruff check: clean

Adversarial review to follow before merge, per the standing gate (MERGE verdict + green CI).

🤖 Generated with Claude Code

[MaxGhenis]

Adversarial review round 1 — verdict: MERGE (risk LOW, zero findings)

Reviewer: gpt-5.6-sol (ultra reasoning), read-only worktree, diff vs origin/codex/thesis-ledger-facts (990fc13 → b6173ba). REPORT verbatim:

REPORT
verdict: MERGE
risk: LOW
findings: []

Highlights of the reviewer's independent verification (full validation record in the review transcript):

• Fold mechanism traced end to end: row deletion + survivor aliases perform the fold; the retire events record the resulting dormant bindings. A retire-only counterfactual regenerated 209 rows (reviving all three UUIDs), and alias-without-deletion was rejected as contradictory curation — confirming all three pieces of the diff are load-bearing.
• Observation conservation proven: 171 distinct current records assigned exactly once before and after; each pair's 1+1 became the survivor's 2.
• Docket safety: all 83 seed names retain identical pre/post candidate UUID sets, all live; the actual pins in the affected families (1ad67747, 8c999902, dbfd5050) sit on separate cadence/entity slices and remain live; no retired UUID or concept appears in the seed.
• Registry: exact 213→216 append, three prior-live same-UUID noted retires, live bindings 209→206 matching the catalog, no UUID recycling.
• Anchor digest recomputed independently (c82492ab…): exactly three identity lines removed, none added.
• Backlog honesty reproduced: the CPI entity-drift fold was undone by regeneration (byte-identical restore), and the initial-claims cadence diagnosis matches the underlying records.
• Test audit: 98 named test functions before and after; none removed or weakened.

Zero findings; merging on MERGE verdict + green CI per the standing gate.