merge: R-B (Crossref references as third citation fallback)

docs/audit/DATA-QUALITY-2026-06-15.md

@@ -382,7 +382,35 @@ is self-contained so a cold session can pick it up.
 - **Acceptance:** the 2 theses gain references; published papers gain
   GROBID-parsed refs beyond OpenAlex/S2 coverage; citing coverage 27/29 → 29/29.
 
-### R-B — Add Crossref as a third citation + abstract source · **MED lever, MED effort**
+### R-B — Add Crossref as a third citation + abstract source · **DONE 2026-06-17**
+
+**Resolution (what was done):**
+- **`crossref.fetch_references(doi)`** added (mirrors `semanticscholar.fetch_references`):
+  GETs `/works/{doi}`, returns one `Citation` per `message.reference[]` entry that
+  carries a registered `DOI` (book/`unstructured`-only refs are skipped — they
+  can't join the graph). `citing_id` is the queried DOI; the ingest supplement
+  rewrites it to `paper.id`.
+- **Wired as the third leg of the citation fallback chain** in `ingest.py`:
+  OpenAlex-empty → S2 → **Crossref** (`_crossref_reference_supplement`, DOI papers
+  only — Crossref's works endpoint is DOI-keyed). cited DOIs pass through
+  `_norm_cited_id` (bare, lower-cased). Network/parse failures degrade to [].
+- **Tests:** `fetch_references` unit tests (DOI edges / `unstructured` skipped / no
+  `reference` key / 404) + ingest tests asserting the Crossref leg fires when
+  OpenAlex+S2 are empty (cited DOI normalised) **and** does *not* fire when OpenAlex
+  has citations. Suite **361 green**, ruff/mypy clean. The DQ-2 abstract half
+  (`fetch_abstract`) already shipped earlier.
+- **Live-validated:** `fetch_references` against the real API returned 33 / 40 / 24
+  DOI-bearing refs for `10.1007/s00454-019-00132-8` / `10.1145/3132705` /
+  `10.1111/cgf.13931` (Springer / ACM / Wiley).
+- **No live corpus change (by design):** as a *fallback*, Crossref fires only when
+  OpenAlex **and** S2 are both empty. Post-DQ-1/R-A the only such papers were the 2
+  `depositonce` theses, which Crossref doesn't index either — so R-B is a
+  forward-looking third leg for future DOI ingests, not a backfill. (Harvesting
+  Crossref's *unique* refs for already-covered papers would need an always-merged
+  supplement instead — a deliberate scope choice not taken, per the doc's "fallback
+  chain" spec.) **Files:** `codex/sources/crossref.py`, `codex/ingest.py`, + tests.
+
+**Original plan (for context):**
 - **What:** new `codex.sources.crossref` client hitting Crossref REST
   `/works/{doi}` for the `reference` array and `abstract`.
 - **Why:** Crossref carries **publisher-deposited references** for many DOIs (free,

docs/infra/grobid-jetson.md

@@ -104,7 +104,36 @@ docker compose logs -f grobid        # wait for "Started ... in N seconds" (mode
 
 The `deploy.resources.limits.memory: 4g` cap in the compose file is honored by
 Compose v2 here (no swarm needed). 3 GB suffices for references; 4 GB leaves
-headroom alongside Postgres.
+headroom alongside Postgres. The `restart: "no"` policy means GROBID does **not**
+auto-start on boot — see "Run on-demand" below.
+
+---
+
+## Run on-demand (stop when idle)
+
+GROBID is only needed **during reference extraction** (`scripts/ra_grobid_backfill.py`);
+nothing else in the stack uses it. Idle it still holds ~3.5 GB, so on the 8 GB Jetson
+keep it stopped and start it only for an R-A run:
+
+```bash
+ssh alfred@192.168.178.103 'docker start papers-grobid'        # ~30s model load
+curl -s --retry 30 --retry-delay 2 --retry-connrefused \
+  http://192.168.178.103:8070/api/isalive                      # wait for: true
+
+# ... run R-A from the Mac ...
+PYTHONPATH=. python scripts/ra_grobid_backfill.py --write
+
+ssh alfred@192.168.178.103 'docker stop papers-grobid'         # reclaim ~3.5 GB
+```
+
+First-time deploy uses `docker compose up -d grobid` (above); afterwards the
+container persists, so `docker start/stop papers-grobid` is all you need.
+
+> Why on-demand rather than a lighter tool: refextract was evaluated head-to-head
+> on the Jetson and rejected — far worse DOI recall on this math corpus (lutz
+> thesis: GROBID **101** usable IDs vs refextract **2**), and it was *slower* per
+> PDF despite a smaller footprint. GROBID stays; running it on-demand removes its
+> only real downside (idle RAM).
 
 ---