Read one run's verdicts in a stable, versioned shape

The automation surface: built for a CI job rather than for a screen. contract_version is the promise -- fields may be ADDED under the same version, but nothing is removed or given a new meaning without it changing, so a client can ignore what it does not recognise and keep working. Attempts paginate on a stable cursor with an explicit has_more; the query count does not grow with the number of attempts. Statuses are a closed set: a run is queued, running, completed, failed, cancelled or budget_exceeded, and an attempt is pass, fail or error. errored_cases is deliberately separate from failed_cases -- an attempt whose synthetic caller or evaluator malfunctioned produced no evidence about the agent, so treating it as a failure would fail a build for a platform outage. voice_evidence_kind says which voice proof an attempt carries: offline_robustness means the authored lines were synthesized and read back, and makes no claim about barge-in, overlapping audio, playout or SIP timing; only realtime evidence can support a timing claim. unproven_assertions counts rules the attempt could not settle for want of that evidence -- they are neither passes nor failures. Transcripts, recordings and rubric rationales are not included; artifacts reports whether each is still retained, expired, or never existed.