Add a citation correctness scorer for RAG outputs

[knisar]

Split from #4 and being implemented in #27 by @M4h1m4. Follows the shape of #12.

A CitationCorrectnessScorer in rai_toolkit/scorers/ that checks whether citations in a response resolve to retrieved context that actually supports the cited claim. This catches two failures GroundednessScorer does not: a citation pointing at a source absent from the context, and a supported claim attributed to the wrong source. A response can be fully grounded and still cite incorrectly.

Conventions agreed on the #4 thread:

• Cited spans are emitted verbatim and verified against the row before being trusted, reusing _verified_evidence_spans as-is (post-fix: accept normalized groundedness evidence spans #18 it handles whitespace/quote normalization with verbatim mapping back to row offsets).
• Keep the existing {response_span, context_span} span shape and carry the citation marker alongside it.
• Behavioral-refusal rows are skipped ahead of the empty-context branch, matching the FactualityJudge ordering and GroundednessScorer should skip behavioral-refusal rows like FactualityJudge #15.
• Rows with no retrieved context return assessed=False, not a defaulted zero.
• A response with no citations returns un-assessed with skipped: "no_citations". Citation presence enforcement is a separate policy metric, out of scope here.
• Tests mock _call_judge, no key or network needed.

Retrieval relevance and context precision/recall stay on #4.

[M4h1m4]

Hi @knisar, PR is up: #27

It follows the conventions agreed here. Three things go beyond the spec, all explained in the PR description: citations resolve in three tiers rather than two, so a numbering mismatch or a label my parser missed is not reported as a fabrication; a fourth un-assessed reason for when nothing resolves; and a confirmed fabrication clamping the score to 0 in Python. That last one is a deliberate divergence, since no existing scorer overrides a judge score.

Assumptions and known limitations are in the PR description.

Happy to rework any part of it, including the approach.