Fuzzy reference matching for DOI-less refs (citation graph robustness)

[deOliveira-R]

Spotted while smoke-testing GROBID on two papers as preconditions for #16:

• Historical (1982 NSE OCR'd, 28 refs): 0 references have DOIs (pre-DOI era).
• Modern (Raj et al. 2022, Annals of Nuclear Energy, 38 refs): only 1 reference has an extractable DOI.

Both cases break the planned DOI-only matching strategy in #16. Pre-DOI literature is unmatched by definition; modern papers fail because GROBID's reference parser doesn't reliably extract DOI strings from the printed bibliography even when the cited works do have DOIs (Crossref-style enrichment isn't part of processFulltextDocument).

Implication for #16

If we ship #16 with DOI-only matching, the citation graph will be sparse for both ends of the corpus — exactly the regimes where it would be most useful. ORPHEUS-relevant historical literature on collision probabilities, transport approximations, etc. won't link at all.

Sketch of fuzzy fallback

For each parsed reference:

1. Try DOI match first (existing Citation graph: fill cites / cited_by from TEI references #16 path).
2. Fallback: title + first-author-family + year tuple against the local items table.
   • Normalize title: lowercase, strip punctuation, collapse whitespace.
   • Token Jaccard ≥ 0.85 OR fuzz ratio ≥ 90 (whichever cheap match works).
   • Author family-name string-equality (case-insensitive) on first author.
   • Year exact match (or ±1 to account for online-first vs print).
3. Fallback 2: venue + volume + page for refs that have a journal abbreviation but no clean title.
   • Venue normalized via the existing dedup venue-canonicalization (if any).
   • First-page exact match.

Acceptance

• Add citations.fuzzy_match(ref, doi_to_id, title_to_id, ...) that returns an item_id or None.
• Wire it into citations.build as a fallback when DOI lookup misses.
• Confidence on fuzzy-matched edges is < 1.0 (e.g. 0.7) so downstream consumers can filter.
• Tests: positive match for title+author+year, negative for near-misses, idempotency on re-run.
• Re-run on the full library after Citation graph: fill cites / cited_by from TEI references #16 ships and report DOI-match vs fuzzy-match edge counts.

Notes

• GROBID 0.8.1 has consistent per-paper output but the DOI-extraction recall in references is the documented weak point. Crossref reverse-lookup by reference text is an option (POST /works?query.bibliographic=...) but adds an external dependency and quota.
• The TEI parser bug fixed in OCR fallback via kind='ocr_text' artifact (concrete path for #13 OCR item) #18-comment thread today (titles falling back to monogr/title when analytic/title is empty) makes title-based fuzzy matching meaningfully more robust on historical references.