Extend evidence-fetch tooling to the source-agnostic policy (multi-source + ephemeral full-text)

[Sanjan-Prabu]

Context. Sourcing policy is now source-agnostic (see #59, AGENTS.md §4): evidence may come from any connected source that asserts the mechanism. But the current fetch wrapper scripts/pubmed_fetch.py only covers PubMed (NCBI / Europe PMC / PubTator3). To honor the policy, the agent needs a unified fetch-and-cache layer over the other sources (e.g. bioRxiv/medRxiv, ChEMBL, ClinicalTrials, DrugBank Mechanism-of-Action), each producing a cached snippet + citation metadata that Layer 4 can verbatim-verify.

Goal. One evidence-fetch layer that supports all sanctioned sources with the same verbatim-snippet + ephemeral-full-text contract.

What to do.

• Add per-source fetchers behind the existing pubmed_fetch.py interface (search / fetch / probe / strip-fulltext), writing the same references_cache/ shape (snippet-able text + metadata).
• Enforce the ephemeral full-text rule uniformly: any non-abstract body is deleted after the record passes QC (keep snippet + citation only) — anti-infringement.
• Keep Layer 4 verification source-agnostic (verbatim substring of whatever was fetched + cached).

Done when. A curation can cite + verbatim-verify evidence from a non-PubMed source, and its full-text body is deleted post-QC.

[Sanjan-Prabu]

Built + verified — source-agnostic evidence-fetch layer. Committed locally (not pushed): evidence layer d3f032e + schema broadening 31eaa61. Left open for your review (DrugBank live-pending; CC-BY-NC policy → #60).

What was built: scripts/evidence_sources/ (per-source fetchers: PubMed adapter, ChEMBL, ClinicalTrials.gov, bioRxiv/medRxiv, DrugBank) behind one unified scripts/evidence_fetch.py CLI (search/fetch/probe/info/strip-fulltext/sources). pubmed_fetch.py kept as the untouched PubMed engine (backward-compat); the PubMed adapter reuses it — one layer, zero PubMed-logic duplication.

Design: QC Layer 4 resolves a reference CURIE to its cache file by a pure string transform (<PREFIX>_<id>.md), so source-agnostic Layer 4 needed only each fetcher to pre-write a matching cache file — no validator logic changed.

Schema change (implements #59): the per-edge reference slot was ^PMID:\d+$, failing non-PubMed refs at Layer 1. Broadened to an EvidenceReference type (PMID/DOI/ChEMBL/bioRxiv/medRxiv/clinicaltrials/DrugBank), additive. Verified: full-corpus Layer-1 PASS 4846/4846 (no regression); ChEMBL ref now passes Layer 1; PMID still passes.

Verified: Layer 4 verbatim-verifies ChEMBL/ClinicalTrials/bioRxiv (live); fabricated-snippet negative control fails; ephemeral deletion works (bioRxiv body 45KB→2.7KB post-QC; full-text snippet stops verifying, abstract still does); backward-compat suite 71 pass/7 skip; new tests 17 pass/3 network-skipped.

Follow-ups: (1) DrugBank fixture-tested only (MyChem lacks mechanism_of_action; go.drugbank.com 403) — needs a licensed access path for live. (2) DrugBank CC-BY-NC commit-vs-ephemeral policy → #60. (3) AGENTS.md §5 / dmdb-references skill should call evidence_fetch.py strip-fulltext --all (all sources) — minor doc follow-up.