feat: add NER entity extraction and knowledge graph to docsaf

[ajroetker]

Summary

• Adds GliNER2-based named entity recognition to the docsaf documentation sync tool via termite's /api/recognize endpoint
• Extracted entities are stored as document metadata (faceted keyword search) and as nodes in a knowledge graph index (entity-based document traversal)
• Graph index uses field-based mentions_entity edge type that automatically creates edges from doc sections to referenced entity nodes

New flags (prepare/sync)

• --ner-model — Termite recognizer model (e.g., fastino/gliner2-base-v1)
• --ner-label — Zero-shot entity labels (repeatable)
• --ner-threshold — Confidence threshold (default 0.5)
• --ner-batch-size — Texts per NER batch (default 32)
• --termite-url — Termite API URL (default http://localhost:8088)

Usage

docsaf sync --dir ./docs --table docs --create-table \
  --ner-model fastino/gliner2-base-v1 \
  --ner-label technology --ner-label concept --ner-label api_endpoint

Test plan

• Build docsaf: cd examples/docsaf && GOWORK=off go build ./...
• Run without NER flags (backward compatible): docsaf sync --dir ./docs --table docs --create-table
• Run with NER: docsaf sync --dir ./docs --table docs --create-table --ner-model fastino/gliner2-base-v1 --ner-label technology --ner-label concept
• Verify entity records in prepared JSON: docsaf prepare --dir ./docs --ner-model fastino/gliner2-base-v1 --ner-label technology && jq 'to_entries[] | select(.value._type == "entity")' docs.json
• Verify graph edges created: query the knowledge graph index for mentions_entity edges
• Verify --ner-model without --ner-label gives a clear error
• Verify load auto-detects entity records and creates graph index