feat(source): the first entitled-network publisher attempt — Cambridge Core book chapters, measured end to end on a subscribing network

[sotashimozono]

doiget 0.8.9 (MCP binary) and 0.8.12 (CLI), measured 2026-09-02 from a subscribing university network.

#517 decided "change it" — doiget retrieves through routes the user is lawfully entitled to use, including the institutional entitlement of the network it is sitting on — and recorded a three-step order: (1) give the closed-DOI path a trace, (2) record the network in provenance, (3) then the publisher attempt. (1) is #505. (2) and (3) have no issue. This is the issue for (3), with a first target that was measured rather than guessed.

The measurement

Target: Chaikin & Lubensky, Principles of Condensed Matter Physics (CUP, 1995). Book DOI 10.1017/CBO9780511813467, chapter DOIs 10.1017/CBO9780511813467.00N.

step	result
fetch 10.1017/CBO9780511813467	oa_status: closed, pdf.status: no_oa_url, type: monograph
fetch 10.1017/CBO9780511813467.006	oa_status: closed, pdf.status: no_oa_url, type: book-chapter, page: 213-287
attempts[] on both	five not_consulted_not_needed rows, no publisher row — the #505 silence
dry_run pdf_sources	one entry, key oa-publisher, 24 hosts, no cambridge.org
landing page, plain curl + UA, university IP	HTTP 200; title="You have access to this product" appears 30 times
chapter PDF from the landing-page href	HTTP 200, application/pdf, 4,102,500 bytes

The href pattern on the landing page is "/core/services/aop-cambridge-core/content/view/<HASH>/<isbn><chap>_p<first>-<last>_CBO.pdf/<slug>.pdf", one per chapter plus frontmatter / contents / preface / glossary / index. Fifteen of them cover pages i-699 contiguously. Text layer present (pdftotext extracts body prose).

Why the Crossref link[] URL is not the route

Crossref carries exactly one link for the chapter:

"link": [{"URL": ".../content/view/7FCF854BE4B7BD006C95BB0E2BBAA209",
          "content-type": "unspecified",
          "intended-application": "similarity-checking"}]

That URL returns the same 4,102,500 bytes. It must still not be used: #517 recorded that the intended-application filter stays, and orchestrator.rs::extract_crossref_publisher_url implements it (ADR-0052). The landing-page href is the same bytes reached the way an entitled reader reaches them, and it is the route with a defensible provenance story. Worth stating in the eventual ADR, because the two URLs being byte-identical is exactly the coincidence that invites the shortcut.

Note also that 0.8.9 still surfaces that Similarity Check URL as oa_url from doiget_resolve_paper; 0.8.12 returns null. Separate issue.

Why CUP book chapters are the right first target

• IP-authenticated, no cookie and no SSO. decision: the fetch path never contacts the publisher for a closed DOI, even from an entitled network — and CONFIG.md §6.1 names two blockers when there are three #517 recorded that natural resolution works for IP entitlement and not for cookie/SSO login. This case is the clean half.
• No bot wall. A plain curl with a browser UA got 200 and a PDF. Contrast ADR-0039 / ADR proposal + evidence: on a subscribing university network, paywalled fetches fail for two independent reasons (ADR-0027 scope, and a publisher WAF that answers scripted clients with 202+empty) #407, where the measured failure was a WAF answering a scripted client with 202 + empty body. So this target exercises the entitlement path without immediately colliding with the WAF problem.
• Segment currently served at zero. type: book-chapter is a class doiget has no route to at all, OA or not.
• Stable, parseable landing page. No JS execution needed; the hrefs are in the served HTML.

Sketch

• Capability gate parallel to Tier 3, but with no API key, because the credential is the network: an explicit opt-in plus a recorded note naming the entitlement. CapabilityProfile (lib.rs:1362) grows one field; the AgreedButNoKey / KeyButNotAgreed pair from CapabilityError does not transfer unchanged and needs its own shape.
• can_serve: prefix 10.1017, type == "book-chapter", grant present.
• A separate allowlist key. Not a widening of oa-publisher — a closed host in the OA list is how a subscription PDF ends up stored as if it were free.
• The store currently means "PDF present = free to use". A subscription PDF stored with license = "unknown" is indistinguishable from an OA one to bib / csl / text. It needs access = "subscription" and the terms URL (https://www.cambridge.org/core/terms) on the entry, and the entitlement recorded per step (2).
• Entitlement check must be positive, not by absence of error: require Content-Type: application/pdf, since an unentitled IP gets 200 + HTML.

The terms guard belongs in the code

CUP's terms permit reading under the subscription and prohibit systematic download of an entire book. That distinction is currently the user's to remember; it should be the source's to enforce:

• refuse book-level DOIs (type: monograph) — chapters only;
• cap per container-title within a session, and say why on refusal.

Filing this as part of the design rather than as a later hardening, because a source that will happily walk fifteen chapter DOIs is the one that makes the tool a liability.

Order

Depends on #505 for (1) and on the provenance schema migration for (2) — per #517, so that the before/after is diffable in the log rather than asserted.

Refs #517, #505, #497, #407, ADR-0019, ADR-0039, ADR-0052.