A discovery campaign: each stage is a Stage-Graph sub-DAG running the active-learning loop until
exit criteria pass. Human gates sit after every stage (sandbox paused at each). Failure loops
return to earlier stages with the failing evidence attached — never a silent restart.