Fold every spelling of a registry the census already counts

[realmarcin]

Refs #84 (the spelling half; the inclusion decision stays open there). Closes #255, closes #256, closes #270, closes #271, closes #272, closes #278, closes #279, closes #280, closes #281, closes #282, closes #283, closes #284, closes #285.

The rule

The census counted each registry under one spelling and dropped the rest. prefix_census.norm now folds every namespace named for, or qualified by, a counted registry into that registry (#271). That covers:

• case and alternate names: gold: and GOLD:; SwissProt:/Swiss: and UniProt:; TAXON: and NCBITaxon:; CAS-RN: and CAS:; TC: and TCDB:; MSH: and MESH:;
• entity-qualified namespaces: kegg.compound:, mediadive.medium:, gtdb.genome:, uniprot.location:, RHEA-COMP:.

gold.ecosystem and pubchem.compound were always folded this way. Reference and curator collections named for a registry are not its terms and stay out: GO_REF: and PO_REF: are literature-like (#84 lists GO_REF as a CITATION candidate), and GOC: is curator attribution, which #84 proposes never to count. Mechs mix lowercase bioregistry spellings with upper-case ones, TaxonMech most heavily.

The list was measured by scanning every census record at the #120 pins with a token pattern allowing dots, underscores and hyphens. An earlier scan without hyphens missed CAS-RN (#270). A spelling a Mech adopts later won't be caught until someone scans again. Each fold was sampled in the pinned records to confirm it is the same registry.

spellings	folds into	occurrences
gold	GOLD	1,492,983
bacdive	BacDive	556,160
gtdb.genome	GTDB	135,164
img.taxon	IMG	68,303
DSM	DSMZ	31,553
pdb.ligand, pdb-ccd, RCSB_PDB, PDBe, pdbe	PDB	15,209
RHEA-COMP	RHEA	11,169
mediadive.medium / .solution / .ingredient	MediaDive	5,347
kegg.compound, kegg.drug, kegg.glycan, kegg.module, KEGG_REACTION, KEGG_PATHWAY	KEGG	3,234
pubchem, pubchem.aid, PubChem_Compound	PubChem	3,291
komodo.medium	KOMODO	3,262
uniprot.location, uniprot.ptm, UniProtKB-KW, SwissProt, swissprot, Swissprot, Swiss, UNIPROT	UniProt	1,290
CAS-RN	CAS	736
chembl	ChEMBL	456
ec	EC	214
ChEBI	CHEBI	177
PubMed, PUBMED	PMID	25
TAXON	NCBITaxon	5
MSH, MeSH	MESH	15
interpro	InterPro	2
TC	TCDB	194

About 66 of the CAS-RN values are EC-list numbers that MediaIngredientMech mislabelled as CAS upstream.

Overlaps

build_subsets.py folds the same spellings for heatmap columns, so record lists agree with the counts.

• Folds that create shared terms: kegg.compound:C00067 is the same term as KEGG:C00067.
• Separate term keys (TERM_SPACE): Rhea compound ids share values with Rhea reaction ids (966 of 2,243), and PDB ligand codes are a different kind of identifier from entries. Both keep their own term keys, so an overlap never pairs a compound with a reaction or shows a ligand as an entry. They still count toward the RHEA and PDB columns and the edge's vocabulary breakdown. The page's term filter matches them under their column too, so filtering an edge by "PDB:" lists the ligand codes (Edge panel's column filter pre-fills 'PDB:', which hides PDB-CCD ligand terms #284).

Review history: five internal review rounds, then a Codex review, which found MSH: (#285). That is folded here too.

Rerun

prefix_census.py and build_subsets.py were rerun at the #120 pins from the same snapshot, followed by build_data.py and assemble_page.py. Against main:

• Unchanged: revisions and record counts. The vocabulary tile goes from 46 to 48: ChEMBL and KOMODO had only ever been written the uncounted way.
• Heatmap:
  • TaxonMech GTDB 648,112 → 783,276;
  • ProteinTraitsMech PDB 235,957 → 250,830, RHEA 613,456 → 624,625, KEGG 245 → 2,018, UniProt +634, CHEBI +172;
  • CellStructureMech UniProt 337 → 993 and PDB 277 → 280;
  • AntibioticMech KEGG 0 → 1,461 and PDB 6 → 339;
  • MediaIngredientMech CAS 727 → 1,463;
  • a few NCBITaxon, InterPro and PMID units in ProteinTraitsMech, and CHEBI +4 in MediaIngredientMech and +1 in CellStructureMech;
  • ProteinTraitsMech TCDB 4,705 → 4,899, in the census only, since TCDB has no column.
  • KEGG moves up the column order, because another Mech now cites it.
• Clickable cells (records per cell):
  • new: AntibioticMech KEGG, 1,154;
  • AntibioticMech PDB 5 → 331;
  • MediaIngredientMech CAS 287 → 739;
  • CellStructureMech UniProt 65 → 184 and PDB 19 → 20;
  • ProteinTraitsMech KEGG 231 → 1,967, UniProt +595, PDB +456, CHEBI +168, NCBITaxon +2;
  • TaxonMech GTDB +447.
• Overlap chords: no Mech pair gains or loses an edge. Two pairs gain an "other vocabularies" chord, because their new shared terms are in vocabularies outside the page's short list.
  • ProteinTraitsMech–AntibioticMech gains 140 shared PDB ligand codes.
  • AntibioticMech–MediaIngredientMech gains 21 CAS numbers.
  • Seven more gain one to five terms each.
• Census date: the note's date moves to 25 September 2026. The census stamps its run date, and that is also the pins' UTC date.
• Checks: tests pass, and assemble_page.py --check reports current. PrefixListTests now unescapes hyphens and checks that every fold is reachable (PrefixListTests' literal_prefix() leaves escaped hyphens in, so its emitted-names set is wrong #280).

Terms from free text (#254, #272). Overlaps read terms from free text as well as from structured fields, for every prefix, and some of the new terms come from prose:

• CHEBI:15982 on the CommunityMech–MediaIngredientMech and ProteinTraitsMech–MediaIngredientMech chords comes from a MediaIngredientMech note that rejects that id;
• the TAXON fold lists ELM:ELME000348 as a Fungi record, though its definition excludes Fungi;
• the 21 CAS terms come from MediaIngredientMech change-log lines.

This predates the PR, applies to all prefixes, and is filed as #254 for its own methodology change.

Not in this PR

Which namespaces the census should count at all (CATH, CDD, PROSITE, StrainInfo, LPSN, TOGO, UNII and about a hundred more) is the maintainer's decision on #84, with the measured inventory in the latest comment there. _fleet/README.md says so, states the fold rule, and corrects its old claim that adding a census prefix adds a heatmap column.

🤖 Generated with Claude Code