feat(nvidia): export Auto Ontology terms as an Ossie ontology and mapping

[liofnv]

Summary

Builds on #589. Please review only the last commit (4e9d27d); the first commit is #589, and this diff shrinks to that commit once #589 merges.

Auto Ontology's terms are business concepts, its attributes are properties of those concepts, and its foreign keys link one concept to another. The converter wrote only the semantic model, so an exported Auto Ontology model showed no concepts, relationships or mappings. The import command can now also write an Ossie ontology and a mapping document that populates it from the semantic model:

uv run ossie-nvidia-auto-ontology import \
  --input tpch.auto_ontology.yaml \
  --output tpch.semantic_model.yaml \
  --ontology-output tpch.ontology.yaml \
  --mapping-output tpch.mapping.yaml

The same is available as convert_auto_ontology_to_ossie_ontology. The mapping references both other documents by name and, from the CLI, by a path relative to itself.

How the ontology is read off the semantic model

Semantic model	Ontology	Mapping
Dataset	Entity concept, named in PascalCase	One concept mapping
Primary key, else the first unique key	identify_by	Object mappings: one expression for a single-field key, referent mappings otherwise
Field with a portable datatype	ManyToOne relationship to the built-in value type, or OneToOne when the field alone identifies the concept	Link mapping from the field
Relationship	ManyToOne relationship to the referenced concept	Link mapping to the referenced concept's objects

A key made of a link and a field, such as a TPC-H line item's order and line number, is mapped through nested referent mappings, as in the specification's own OrderLineItem example.

Nothing is guessed:

• A field whose type is unknown, or has no built-in value type (Time, Opaque), is left out.
• A concept whose dataset has no usable key is declared but not mapped.
• Relationships are verbalized from their field names ({Customer} has Account Balance {Decimal}), because Auto Ontology records no verbalizations.

Export changes the mapping needs

• Names. Auto Ontology keeps term names unique across every database it catalogs, so it renames a term to Name (database.schema.table) when another database already uses the name. An Ossie document is its own namespace, so the dataset is exported as Name wherever that is free in the document. The term keeps its full name on the way back.
• Key columns. Auto Ontology gives a foreign-key column no attribute. A key or relationship column that no field exposes is now exported as a field, so the mapping can reference it. It is marked so it does not become an attribute on the way back.

Testing

• uv run pytest in converters/nvidia: 118 passed, including 13 new tests (10 in tests/test_ontology.py).
• The three documents the CLI writes pass validation/validate.py against the core, ontology and mapping schemas (test_cli_writes_three_documents_that_pass_official_validation).
• The generated ontology parses with converters/ontology's OssieParser.
• A live export of SNOWFLAKE_SAMPLE_DATA.TPCH_SF1 gives 8 concepts and 64 relationships. 6 of the 8 concepts are mapped: LINEITEM and PARTSUPP have no key, because Snowflake declares none and their keys are composite. With the TPC-H primary keys declared, all 8 are mapped.

AI disclosure

This contribution was prepared with AI assistance (Claude Code), per the ASF Generative Tooling Guidance.

[liofnv]

@jbonofre

[jbonofre]

Thanks @liofnv for this PR!

The ontology and standalone mapping report (ontology.py) is very clean, well-aligned with the ontology/mapping specifications, and the round-trip handling for database-qualified term names and key_column fields is well thought out. Great test coverage with the schema validator as well.

Two minor observations for potential follow-up:

1. In _identity() (ontology.py), checking set(relationship["from_columns]") <= set(key) in addition to column in relationship["from_columns"] would prevent a composite foreign key from being treated as an identifying link when only a subset of its columns is part of the dataset's key.
2. In _referent_mappings(), tracking a visited set of concepts when recursing through identifying link would guard against RecursionError if a schema ever has cyclic identifying relationships.

LGTM 👍

Let's merge #589 first once CI completed, then I will merge this one.