See what a batch import actually cost, and export it

A batch execution now reports how it read: records and pages, throughput, how much of the working time went on waiting for the source, megabytes transferred, and — per step — the split between remote wait and local work, page latency at p50, p95 and max, and the cost of a page plotted against how far the scan has got. A flat line means cost is per-record work; a climbing one means every page is paying again for the records behind it. The settings that produced those numbers are printed beside them, so a measurement is never separated from what it measured.

Because the figures change while a run is going and the page keeps no history, they export. A Markdown report carries the totals and per-step tables for a ticket or a pull request, describing each curve in words since it has no chart. An HTML report carries the charts drawn, as a single file that opens in any browser with nothing fetched from the network, and hands the plotted curve back as CSV for a spreadsheet or for comparing two runs point by point. The page itself prints.

Reach it by adding /metrics to a job's address. It is out of the navigation on purpose: it answers why a run was slow, not whether it worked.