This ad-hoc case expects the agent to:

- compare the latest completed Beijing `10:50-11:00` and `11:00-11:10`
  windows for `the primary premium cluster`
- use only `clinic_data_proxy.slow_query_logs`
- aggregate slow queries by `digest`, rank by occurrence count, and take the
  top 10 in each window
- identify digests present in the current-window top 10 but absent from the
  previous-window top 10
- return only a structured JSON payload with:
  - `cluster_id`
  - `comparison`
  - `evidence_sufficient`
  - `new_sql_digests[]` containing `sql_digest`, `sql_statement`, and
    `occurrence_count`