NU-008 Compiler: SchemaLLMPathExtractor with provenance enrichment #22

Part of #30.

The Compiler: turn an approved Reviewed Source into typed graph triples with an explicit
decisions[].

Build

SchemaLLMPathExtractor with:

possible_entities  = Literal["DECISION", "ENTITY", "EVIDENCE"]
possible_relations = Literal["SUPERSEDES", "EVIDENCED_BY", "AFFECTS", "DEPENDS_ON", "PART_OF"]
kg_validation_schema = {...}
strict = True

Pydantic-validated triples — the schema guardrail is enforced by the library rather than by prompt
discipline.

LLM is OpenAILike so any OpenAI-compatible endpoint works. Add a startup check that structured
output actually works on the configured endpoint — compatible endpoints vary wildly in JSON-mode
fidelity, and SchemaLLMPathExtractor depends on it. Fail loudly at boot, not at 3am on malformed
triples.

The write-side hole this closes

PRD §3 defines LLM-Wiki as a Markdown document with four sections — Executive Summary, Core
Entities & Relationships, Known Issues & Verified Solutions, Cross-References — and calls it "the
only form the graph ever ingests." None of those four sections is a Decisions section, yet
FR-5 requires persisting Decisions out of it. Worse, author — which FR-5 makes mandatory on
every Decision node — appears in neither upstream schema.

So LLM-Wiki becomes structured output with an explicit decisions[], not prose. If it stayed
prose, author/timestamp/evidence_span would exist only as words inside a paragraph and
persistence would have to re-extract them — a second lossy pass, and two places that can
disagree about what the author's name was.

Enrichment seam the library does not cover

SchemaLLMPathExtractor yields typed nodes and edges. It does not carry author,
author_source, timestamp, evidence_span, or provenance ref-sets. allowed_entity_props is a
list of permitted property names, not a typed validated model — and author_source originates
in the parser, which the extractor never sees. Needs a custom extractor subclass or a
post-extraction enrichment step stamping these before insert_nodes().

Do not use PropertyGraphIndex.from_documents() — it does chunk→extract→persist in one shot,
and this pipeline splits those across two human gates.