███    ███ ███████ ████████  █████  ██████   ██████  ██
████  ████ ██         ██    ██   ██ ██   ██ ██    ██ ██
██ ████ ██ █████      ██    ███████ ██   ██ ██    ██ ██
██  ██  ██ ██         ██    ██   ██ ██   ██ ██    ██ ██
██      ██ ███████    ██    ██   ██ ██████   ██████  ██

── Data sources ─────────────────────────────────────────────────────────────────
- Crossref (crossref.org), DataCite (datacite.org), OpenAlex (openalex.org)
  Primary metadata (title, abstract) and DOI resolution.
- Retraction Watch (gitlab.com/crossref/retraction-watch-data)
  Title, status, and retraction reasons. Synced daily.
  Reinstatements act as chronological resets; retractions supersede expressions
  of concern, pooling reasons exclusively from the dominant status (dedup).
- Unpaywall (unpaywall.org)
  Open-access routing with fallback to native doi.org resolution.

── Abstract extraction mechanism ────────────────────────────────────────────────
Greedy facility-location coverage using BGE-M3 (BAAI/bge-m3) sentence embeddings.
First pick: the centroid-nearest, semantically representative sentence.
Subsequent picks: max-marginal gain until gain ≤ μ (mean pairwise similarity).
Output: selected sentences in original order, with "…" marking gaps.