fix(rpiv-workflow): hard-fail infra-death stops on collect-all fanout units

A collected soft-halt is a permanent skip — the resume fold rebuilds the
failedOutput sentinel by unitIndex and never re-dispatches the unit, so
collecting a transient infra death (error/noResponse/toolUse: the child
session never delivered a complete pass) turned one OOM'd unit into an
unrepairable hole downstream. Such stops now take the fail-fast terminal
halt even under collectAll: the trail carries an UNCOLLECTED failed unit
row, the slot stays unfilled, and resume re-dispatches exactly the dead
unit. Semantic failures (extraction/validation/timeout/length) stay
collected — there the skill ran and was judged.