Normalize null judge detail lists only for positive verdicts

[lipikaramaswamy]

Problem

LLM judge responses may emit an explicit JSON null for a detail list whose documented empty state is an empty array. Pydantic defaults apply when a field is omitted, but not when it is present as null, so an otherwise usable positive judgment becomes unavailable.

This is related to the model-output drift handled in #268 and the conservative rewrite-evaluator normalization in #269.

Scope

Apply verdict-aware normalization to these evaluation schemas:

• DetectionJudgmentSchema.invalid_entities
• TypeFidelityJudgmentSchema.invalid_replacements
• RelationalConsistencyJudgmentSchema.relations
• AttributeFidelityJudgmentSchema.entities

Required behavior

Normalize null to an empty list only when the associated verdict is positive:

• all_valid=true plus invalid_entities=null becomes invalid_entities=[]
• all_valid=true plus invalid_replacements=null becomes invalid_replacements=[]
• all_consistent=true plus relations=null becomes relations=[]
• all_valid=true plus entities=null becomes entities=[]

Keep negative or inconsistent payloads invalid:

• a negative verdict plus a null or empty detail list must not be accepted as a complete judgment
• malformed output must continue to surface as judge unavailable rather than fabricating a verdict

Do not normalize nested RelationCheck.entities or AttributeCheck.attributes_checked; an emitted check requires those details.

Acceptance criteria

• Add model-level validation implementing the positive-verdict-only rule for all four schemas.
• Add parameterized tests covering positive/null, positive/empty, negative/null, negative/empty, and valid negative/non-empty cases.
• Preserve the current postprocessing behavior for malformed judgments.
• No public API changes.
• make format-check and make typecheck pass.