Before re-running a validation on a new dataset, regenerate derived inputs (name maps, configs) with the current script version

When a check pipeline evolves during a session, any pre-generated inputs for later datasets go stale silently. We had made a signal name map for a second, larger dump early on. Later we fixed the generator script for the first dump, adding an alias for array-element names; without the alias, ~13k signals showed as 'missing'. The second dump's pre-made map still came from the old script. A cheap habit catches it: before launching the long job on the next dataset, regenerate every derived input with the current scripts and cmp it against the old copy. If it differs, use the new one and say so in the method notes. Here it differed, which would have cost a ~1.5 h job plus a confusing 'missing' count. Also: when a multi-hour streaming conversion (A | B) finishes, record both pipe stages' exit codes (bash PIPESTATUS), not just the last one. A producer can fail while the consumer exits 0 on a truncated stream.