Reclassify loose/historical/lossy relationship types out of `Identity` in `predicate_mapping.csv`

[nicoloesch]

Is there an existing issue for this?

• I have searched the existing issues

Bug summary

GroundingConstraints trusts PredicateKind.IDENTITY edges by default for hierarchy-anchor pathfinding (find_standard_paths), treating them as safe "same concept" hops. Several relationship_ids in config/predicate_mapping.csv were classified under Identity despite being loose, historical, exception, or class-to-instance relations rather than true 1:1 equivalences. This let find_standard_paths walk through unrelated concepts and surface spuriously perfect (relevance == 1.0) matches with no real relationship to the query.

Root cause

config/predicate_mapping.csv defines, per OMOP relationship_id, a (class, subclass) pair loaded into relationship_mapping and used as the predicate_kind filter. A full audit of all 735 rows (all five classes) found 27 rows misclassified under Identity (or left entirely unclassified) that should not be trusted as identity hops:

• SNOMED historical-association family: Concept poss_eq from/to (possibly_equivalent_to), Concept was_a from/to (was_a), Concept alt_to from/to (alternative_to). These are SNOMED's own ambiguous/fallback inactivation relations, explicitly documented as requiring human judgment to resolve and are not safe for automated traversal.
• Cross-vocabulary class-to-instance mappings: OPCS4 - SNOMED/SNOMED - OPCS4 (own name says "equivalent or categorical"), ATC - RxNorm/RxNorm - ATC, ETC - RxNorm, ICD9CM - MedDRA/MedDRA - ICD9CM, ICDO - SNOMED/SNOMED - ICDO. These connect a class/category-level concept to a specific instance (e.g. one row's own example shows a 7-ingredient combination product mapping to the ATC class of just one of its ingredients).
• Structural/substance-variant relations mislabeled as lifecycle events: Modification of/Has modification, Reformulated in/Reformulation of. These are salt/ester/formulation variants, not deprecated-concept links.
• OMOP-internal mapping-exception bookkeeping: Excluded in map from/Included in map from/Map excludes child/Map includes child. These describe exceptions to a mapping rule, not equivalence between two concepts.
• A grouping construct, not an equivalence: Proc Schema to ICDO. Its own example links a cancer-registry anatomic schema to one specific malignancy occurring at that site.
• Previously unclassified rows: Reformulated in/Reformulation of, Has inherent, Has transformation/Transformation of were blank (class/subclass), flagged "left for manual review," and silently dropped from relationship_mapping entirely. Classified for the first time, by precedent from sibling rows already in the file.

Fix

• Reclassify the 27 affected rows in config/predicate_mapping.csv to honest, lower-trust classes (Association,Administrative, Association,Mapping, Hierarchy,Taxonomic – up, Composition,Structural, Composition,Logical, Attribute,Context, Attribute,Pharmaceutical).
• Add one new subclass, Association,Mapping, for "cross-vocabulary or historical mapping without a guaranteed 1:1 correspondence". There isn't an existing subclass outside Identity that fit this shape.

Code for reproduction

## Evidence

> Query *"Galactosaemia"* → target *Galactosemia* (439788). At a broader hierarchy anchor (level 3), the top-scoring result became **Cystinuria** (4225406), a completely unrelated amino-acid transport disorder, with `relevance == 1.0` and no resolver attribution.
>
> Traced path: `Galactosemia (439788) -[possibly_equivalent_to]-> 40351277 (a retired SNOMED "disjunction" concept literally named as a list of ~6 unrelated diseases) -[possibly_equivalent_to]-> Cystinuria (4225406)`.

`possibly_equivalent_to` is SNOMED's documented many-to-many fallback for retired concepts with no clean 1:1 successor. This is the opposite of "Exact equivalence," which is how it was classified.

Error messages