List the slowest marketing runs whose run falls in week_start's week.

Backs the click-through on the weekly run-latency chart, and answers "which runs took really long, and was it waiting or working?" -- every row breaks end-to-end latency into queue wait and the prep-plus-analysis span, because the same ~40-minute total is a dispatch problem in one run and slow work in another, and the total alone cannot tell them apart.

Uses the same week bucketing (date_trunc('week', run_time)) and the same _latency_filters population as _run_latency_weekly_series, so total equals the clicked point's runs for the same scope -- and the first row under the default sort is the run behind that week's end_to_end_max_seconds. The same survivorship caveat applies: a run that failed or is still stuck has no analysis_completed_at and cannot appear here, however long it has been waiting. Each row carries file_id to deeplink into the file's history view.

Paginated with limit/offset over the full ordered set. The sort key is followed by run_id as a unique tiebreaker so equal latencies cannot shuffle between pages, which would otherwise let a row appear twice or not at all. total is always the whole filtered set, not the page.