fix: handle string attributes in graph ontology to prevent TypeError crash

[octo-patch]

Fixes #135

Problem

When the LLM returns ontology attributes as a plain list of strings (e.g. ["full_name", "role"]) instead of the expected list of dicts (e.g. [{"name": "full_name", "type": "text", "description": "..."}]), set_ontology() in graph_builder.py crashes with:

TypeError: string indices must be integers, not 'str'

at attr_def["name"] for both entity and edge attribute loops.

Solution

Two-layer fix:

1. ontology_generator.py (primary fix): Normalize string attributes to properly-structured dicts during _validate_ontology_result(). Any attribute that arrives as a bare string s is converted to {"name": s, "type": "text", "description": s} before it reaches downstream code.

2. graph_builder.py (safety net): Add isinstance(attr_def, str) guard in both the entity and edge attribute loops inside set_ontology(). This ensures the crash cannot occur even if the ontology was built through a path that skips the generator's validation.

Testing

Tested with an ontology payload where attributes is a list of strings — graph build now completes without error. Existing dict-format attributes continue to work as before.