Pre-qualify the existing render-job backlog by cancelling jobs whose source article body falls below min_words. PR #184 added an intake- side gate, but the queue still holds thousands of pre-gate stubs that would render into 8-second clickbait. This endpoint sweeps them in bounded batches; call repeatedly with dry_run=false until would_cancel drops to zero.

Body: min_words: int (optional; default INTAKE_MIN_WORDS env, then 500) dry_run: bool (default true) limit: int (default 5000) — max render_jobs rows examined per call status_filter: list[str] (default ["queued", "deferred"])

Jobs with a NULL article_id (e.g. forwarded-newsletter chunks with no 1:1 article row) are skipped — we only judge jobs we can word-count.