D2 component 4: LLM graded-relevance judge. For each (query, candidate) pair it asks a complete_json_schema/4-capable LLM (default: z.ai GLM-5.2) for a 0..3 relevance grade, and emits source: :llm label attrs recording the judge model.

Integrity contract: unlike the entity extractor, there is NO safe fallback grade — a fabricated default would bias D2's optimization. So on an incapable/absent LLM, any error, an unparseable result, or a grade outside 0..3, the pair is SKIPPED and reported under :errors (with its query_ref/memory_ref and a reason), never labeled. The bulk pass therefore reports exactly what it graded and what it could not.

The pair's split is preserved on the emitted label so a :calibration pair's llm row lines up with its human row for the Cohen's-kappa join. Structure + parsing are pure and fake-LLM testable; only the live call is paid.