Legacy ontology payloads could be saved with attributes as free-form maps of attribute names to descriptions. The graph build path dereferenced attr_def["name"] directly, which raised a KeyError before the Zep ontology request. This adds a shared ontology normalizer, applies it when validating generated ontologies and when setting graph ontology, and adds regression tests for both normalization and graph-builder compatibility.
Constraint: The configured OpenAI-compatible provider guarantees valid JSON but not strict schema adherence for ontology attributes
Rejected: Guard only graph_builder.set_ontology against missing keys | would leave malformed ontologies persisted and unvalidated upstream
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep ontology payloads normalized at service boundaries so new generations and saved legacy projects share one schema contract
Tested: Targeted backend pytest for ontology normalization in the docker compose runtime
Tested: Python compileall for backend app and tests in the docker compose runtime
Tested: Reproduced the original saved project payload through graph_builder.set_ontology with a stubbed client
Not-tested: Live Zep API round-trip against the upstream service
Jeffery