Iterates every IndicatorSummary in the DB and asks the reviewer agent
(services/output_review.review_read) whether each row's content is
publishable. Rows the reviewer flags as unclean are deleted along
with their translation rows. The API's existing fallback path —
serve the latest IndicatorSummary by (group, tone) — picks up the
previous clean row automatically.
Concurrency defaults to 8 reviewer calls in flight; on the 3245-row
prod archive that completes in ~10 minutes for ~$1 of Haiku cost.
Idempotent: a second run only re-evaluates whatever's still in the
table. --dry-run skips the deletion stage. After the live pipeline
fix landed (JSON-mode + reviewer at write time) this script should
not find anything on subsequent invocations.
