Agent wait loops: one long background wait beats repeated polls

When an agent waits on a long job (e.g. a rate-limited scraper taking an hour), every 10-minute blocking poll is a full model turn that re-reads the conversation; late in a long session each turn cost roughly as much as a substantive step. Start long jobs as early as possible, then do a single background wait with a long timeout (until ! pgrep -f 'job[.]py' >/dev/null; do sleep 30; done) and do useful offline work (drafting the write-up from partial results) while it runs. Restrict the slow job to the cases that affect the headline numbers first (e.g. sort by impact) so it can be stopped early with a clear, documented scope.