ProvenanceBench: scoring and verifier alignment

[aykoooo]

What & why

Refs #36

Define how ProvenanceBench scores detect.py output and how that maps to aibom_verifier verdicts (verified_derivative, fraudulent_claim, incompatible, unsupported, insufficient_evidence).

Scope

• Document gold y vs detector pred (related / unrelated / abstain) and confidence float role.
• Threshold policy: when does a float become confirm vs reject vs abstain?
• Map bench labels to verifier taxonomy (table in issue or wiki).
• Optional: per-signal floats in run log for triage; calibration metrics later (not v1 gate).
• label_confidence: disputed rows excluded from primary bal_acc or scored separately.

Out of scope

• Implementing real weight signals in detect.py (separate work).
• Per-run random negative generation.

Test plan

• Written spec accepted on issue
• Eval harness metrics align with spec (or doc explains gap)

[aykoooo]

Decision: ProvenanceBench v1 methods return Result(verdict, score, diagnostics, note, abstain_reason).

• verdict is related, unrelated, or abstain; score is nullable so an abstention does not imply a fabricated numeric confidence.
• expected=unsupported is scored correct when verdict=abstain.
• Evaluation metrics and error classes are produced only in the review artifact, never the method artifact.
• The schema migration does not claim a new aibom_verifier verdict mapping or add real weight-level signals; those remain separate follow-up work.