[cbizon]

NLM stopped updating the top-level mesh.nt.gz after MeSH 2024, so Babel's MeSH had fallen behind its UMLS: with UMLS 2026AA, 1,878 MSH codes on 2,162 CUIs had no MeSH record, mostly new Supplementary Concept Records such as foliglurax MESH:C000730045 and befotertinib MESH:C000731667. This PR pins the MeSH release with a new mesh_year in config.yaml, beside umls_version, and downloads that year's RDF dump. Per-year RDF has carried no registry numbers since MeSH 2025, so CAS, UNII and NCBITaxon registry numbers now come from the MeSH XML.

So that this can't happen silently again, the build now fails early when mesh_year lags umls_version, and each release archives the UMLS/MeSH skew report and prints its count in the drafted notes.

The calls worth overruling

• The skew limit is 100 missing codes, and it blocks MeSH ownership. rule check_umls_mesh_skew raises when more UMLS MSH codes than umls_mesh_skew_max_missing_codes have no record in the dump. It leaves out the U "(MeSH Category)" headings UMLS adds itself, and the SUPPRESS=O/E rows check_mrconso_line() drops everywhere else (2026AA has none for MSH). Against UMLS 2026AA, MeSH 2026 misses 0, MeSH 2025 829 and MeSH 2024 1,856 (table and script). Stop the same CURIE leading cliques in two compendia, and fail the build when it does #1124's umls_mesh_ownership waits for the check, so the five MeSH-owning pipelines stop before building compendia.
• Registry numbers come from XML, kept to records in the RDF dump. pull_mesh_registry() stream-parses each record's preferred-concept RegistryNumber from desc{year}.gz / supp{year}.gz (new get_mesh_xml rule). The current year's XML is updated daily while the RDF is a snapshot, so on 2026-10-21 it had registry numbers for 100 SCRs mesh2026.nt lacks; only records with a label in MESH/labels are kept, so the unfiltered NCBI_MESH concord can't pull a label-less MeSH ID into a taxon clique. (Filtering it by taxon/ids/MESH instead would also drop existing links.) Against the old RDF registry (measured on 3a96a82), 158,927 pairs match; of the 2,327 only the RDF had, all but 11 are on records NLM has since deleted.
• No fallback. If NLM hasn't published the configured year, get_mesh and get_mesh_xml fail rather than fall back to the stale top-level file. The dump is hashed as it downloads and checked against NLM's .sha1; a network test checks the year is published.
• Per-year IRIs are normalized on download. normalize_mesh_nt() rewrites year-versioned record IRIs (http://id.nlm.nih.gov/mesh/2026/D002839) to the unversioned form, so the existing SPARQL and get_mesh_id_from_iri() work unchanged, and raises if the versioned prefix never appears.

Details: Ingestion.md, Source data and Architecture.md, Build controls.

Also here

• Registry concord metadata names the XML files actually downloaded (recorded in MESH/xml_sources.txt, since current vs. archived directory is only known at download time) and the RDF dump they are filtered to.
• archive_build.py requires reports/umls_mesh_skew.json, and draft_release_notes.py prints its count and reports MeSH version bumps; for the first release after this one, whose baseline has no mesh_year, it drafts a to-do to name the old version by hand.
• config.yaml and the ingestion doc point at https://nlmpubs.nlm.nih.gov/projects/mesh/rdf/ for checking whether a new MeSH year is out. Downloads stay on FTP here; Download MeSH over HTTPS from NLM's nlmpubs site instead of FTP #1161, stacked on this PR, moves them to HTTPS.

What it does not do

Registry numbers on records already in the dump can still change from one day's XML to the next, so two downloads of the current mesh_year may give different concords until NLM archives the year.

Effect on #1124's MeSH ownership

Replaying #1124's rule on ids/MESH rebuilt from MeSH 2026 (output): MeSH 2026 adds about 1,700 records and drops about 4,000 NLM deleted; the rule moves 46 more CUIs than on the 2024 dump, and all 949 resolvable MeSH duplicate leaders are still predicted resolved. The first cluster build confirms this (#1158).

Verification

On 41b6dc2 the unit tests (730-odd) pass and ruff, snakefmt and rumdl are clean; fixtures are copied verbatim from mesh2026.nt, desc2026/supp2026 and 2026AA MRCONSO.RRF, and the download's checksum handling is tested against a mocked FTP server. On 41b6dc2 get_mesh and get_mesh_xml ran through Snakemake against NLM, and pull_mesh_registry() ran on the real XML. On c6a79c8 the full 2026 dump loaded through the existing Mesh queries, and check_umls_mesh_skew passed with 0 missing codes on the local 2026AA and MeSH 2026 downloads and raised on the MeSH 2025 report.

Nothing blocks this PR.

🤖 Generated with Claude Code