EPIC: General extraction and canonicalization

[kmosoti]

Outcome

Non-fixture source spans can produce validated graph candidates through a provider-abstracted LLM boundary.

Scope

• LLM adapter interface
• Extraction Pydantic schemas
• Prompt contracts
• General source span extraction service
• Title normalizer and alias service
• Graph writer
• Extraction audit trail

Source Docs

• docs/01-requirements.md: Extraction, Canonicalization
• docs/07-engineering-standards.md: Protocol Rules, Contracts, Token And LLM Cost Controls

Done When

• LLM outputs validate through schemas.
• Invalid outputs fail safely.
• Extracted nodes/edges retain source span provenance.
• Deterministic transforms are not delegated to the LLM.