collect_ftd extended one list of dicts across all 411 fails-to-deliver archives -- 22.9 million records, some fifteen gigabytes of Python objects, and the job was OOM-killed at that line on 16 GB. Each archive now becomes a frame as it arrives and they are concatenated once; the same records as Arrow are under two gigabytes. Same rows, same order.
