feat: deterministic reference verification for origin archives

[sdgsfh]

Problem

The acceptance gate in run_archive.py (verify_math_object_origin_archive) combines one deterministic layer (format checks) with a model review across three dimensions. The historical dimension currently has no deterministic support: whether a cited source actually exists — and whether its journal, volume, year, and pages are copied correctly — is judged by the model alone. That is precisely where archives fail silently, because a hallucinated or miscopied citation still looks plausible to a reviewer model.

Design

Add tools/verify_references.py, a deterministic reference checker that runs before/around any model review:

• Structure (offline): contiguous [n] numbering in the Sources section; every in-text citation [n] resolves to a listed source; uncited entries are warnings, not failures.
• Existence: each source is resolved against external indexes — Crossref (DOI lookup or scored bibliographic query), OpenAlex, zbMATH Open, the Internet Archive, and Open Library. A confident match requires title similarity and surname overlap and a year within ±2, so a real title with fabricated authors (or vice versa) does not pass. zbMATH Open and the Internet Archive extend coverage to the 19th-century literature (Puiseux 1850, Jordan 1870, …) that origin archives routinely cite.
• Link liveness: DOIs/URLs must resolve (warning by default; --strict escalates to failure).
• Consistency: year/volume/pages claimed by an entry are compared against the matched records. Claims contradicted by every matched record are failures; disagreements between the indexes themselves (e.g., first-page-only records) are reported as warnings, not silently resolved.

The tool is stdlib-only and never judges whether a source supports a claim — claim-level support stays with the model review. Exit codes: 0 pass / 1 fail / 2 usage error. --json emits a machine-readable report; --offline restricts all lookups to a JSON cache so CI runs without network.

Implementation

• tools/verify_references.py — CLI and library (verify_archive_text()); fetchers are injectable for testing.
• tests/test_verify_references.py — 22 offline unittest cases whose fixtures are trimmed real API responses from the five indexes (Keller 1939 via Crossref DOI + OpenAlex + zbMATH, Alon 1999 via bibliographic query, Jordan 1870 via the Internet Archive, Dudeney 1917 via Open Library, …).
• Negative controls: fabricated citation rejected; wrong DOI attached to an unrelated title rejected; pages contradicted by both indexes rejected; non-contiguous numbering and dangling in-text citations rejected.
• README section documenting scope, usage, and the integration boundary.
• No changes to run_archive.py or any production logic.

Verification

• python -m unittest discover -s tests -p "test_*.py" — 39/39 pass (22 new + the existing 17 runner regressions), Python 3.11.
• Live run against 9 real origin archives produced with this pipeline (34 references total): 34/35 sources verified. The single failure is the honest boundary — Riemann's 1857 Göttingen academy paper pre-dates zbMATH's 1868 coverage and is absent from the book indexes — and the tool reports it as unverifiable rather than passing it.
• The live run surfaced three real-world data quirks, each now a regression test: Elsevier DOIs containing parentheses (10.1016/0040-9383(87)90009-7), Crossref family fields that carry initials ("H. Kauffman"), and Crossref titles with footnote suffixes ("…cell complexes††Partially supported by…").

Limitations

• Claim-level evidence support remains with the model reviewer; this tool only guarantees that cited sources exist and are copied correctly.
• Title extraction from free-form citations is heuristic (italic spans preferred; comma-segment analysis otherwise); exotic citation styles may need the documented conventions.
• Book/article title collisions are handled with publisher-hint routing; an entry that matches only a journal article while claiming book-series data will still be flagged.
• Link checks are warning-level by default because some publishers rate-limit HEAD/GET (observed: intermittent 503s from doi.org for AMS journals); --strict is available for hard enforcement.