Citation validation

Check whether the sources the model cited actually appear in the retrieved chunk set, and log a warning when it cites something it was not given. Surface the outcome on the response as citations_valid and unsupported_citations.

This is the smallest possible hallucination detector, and it catches the most damaging failure mode in a support agent - a confident answer attributed to a document that was never retrieved. Everything in Phase 3's guardrail layer is a generalisation of this.