feat(scholar): Phase 18 C2 tail — PMC references, ar5iv full text, Crossref

[bryanmatthewsimonson]

The deferred C2 slice from COMPLEX_CONTENT_DESIGN.md §4.3 — the "independent quick wins" that never landed. Built while you were mid-analysis, so dist/ is deliberately untouched: everything is verified with node --test over src/ directly (1789 pass), and your loaded extension is unaffected until you rebuild.

Directly serves the corpus you're building — PMC and arXiv are where the origins literature lives.

What you get

PMC (pmc.ncbi.nlm.nih.gov, legacy ncbi.nlm.nih.gov/pmc/ too): captures now carry the reference list as structure — year/DOI/PMID per entry, title/authors only when the DOM marks them (never guessed from prose — the parser has an explicit honesty rule: wrong structure is worse than none), figure captions, and PMCID/PMID ids. references is a local capture record per the design's §7 — no wire change, and the natural feed for citation-graph work later.

arXiv (arxiv.org/abs/…): the abs page is abstract-only, so the capture prefers the ar5iv full-text rendition when it's meaningfully longer (≥2× and ≥2000 chars). Identity stays the /abs/ URL; capture_url records the ar5iv address actually fetched — ar5iv is a machine conversion of the LaTeX source, not the PDF of record, which is exactly why the provenance is explicit. Fails open to the abstract capture on any miss.

Crossref (shared/crossref.js + the SW message): validated-DOI lookup that fills only fields the page didn't provide ("the captured text is always what the page said"). The reader-side trigger is deliberately deferred — see below.

What adversarial review caught before you saw this

A 3-lens verify pass with executed repros broke the first cut seven ways; every fix is pinned by a regression test whose fixture is the repro:

• Google Scholar search links landed in entry.url on essentially every modern-PMC reference (the [Google Scholar] decoration anchor). Citation-service chrome is now excluded.
• The year inside an Elsevier DOI beat the true year (10.1016/j.vaccine.2015… → 2015 instead of the citation's 2021). Year-matching now runs over DOI/URL-stripped text.
• SICI DOIs truncated at their literal < and shipped as wrong identifiers — now omitted entirely (absent beats wrong).
• Italicized journal names adopted as titles (APA style italicizes journals). The journal is the segment the year follows directly — that positional check rejects it while a genuine italicized title still lands.
• 'Lovelace, Ada, and Charles Babbage' mis-split into three authors. Single-token comma parts now bail to absent.
• The ar5iv body swap left wordCount/links describing the abstract — and both are wire-bound (word_count and link tags on the 30023). They now re-derive from the adopted body; links come from the same parsed DOM, or go null ("not captured" ≠ "zero links").
• The proposed Crossref reader trigger was hazardous twice: no scholar display exists to refresh, and the naive session-stash write nulls sourceTabId, silently rerouting NIP-07 publishes through the worker signer. Deferred until there's a surface and a read-modify-write; documented in ROADMAP + JOURNAL.

Wiring notes

• The scholar-meta pass is hoisted above handler dispatch — the arXiv handler reads scholar.arxiv_id, and with the old ordering it was a permanent no-op. Safe: the pass is side-effect-free and substack never reads scholar.
• xray:scholar:fetch is host-allowlisted (ar5iv/arxiv only) and xray:scholar:crossref accepts a DOI, never a URL — the SW must not grow a generic fetch proxy one message at a time.
• Detector: arXiv is exact-host (so ar5iv tabs aren't claimed); PMC goes through isPmcPage (so gene/pubmed pages fall through to generic).

Verification

• No wire-format change — no kind, tag, or content field moves; references/pmc/scholar.rendition are local capture fields.
• npm test — 1789 pass, 0 fail (30 new) · node --check on every touched file · lint 0 errors
• npm run build was NOT run — after merging, rebuild + reload the extension whenever you're done with the analysis; nothing here is live until then.

Independent of the archive stack (#185–#187) — based on main, no shared files with those PRs except none.

🤖 Generated with Claude Code