Effort versus prose instructions

This is a factorial experiment with two levers: one sentence of prose and the --effort flag. Everything else stays fixed (packs, schema, user turn, model, three trials). The design decision that matters most is the two new packs: each issue is invisible in any one search, so they measure exactly the behaviour the "cross-check" sentence claims to cause, and recall on them is reported apart from the ordinary action items, which every cell already finds.