Scaling-ladder drivers: don't project memory from the N=1 per-process peak

A concurrency ladder that skips the next N when 'projected memory' exceeds RAM goes wrong if it projects from the largest per-process PSS/RSS seen, which is usually the N=1 value. For mmap-based readers most of that is shared file pages, which split among sharers as N grows. In one run this projected 757 GiB for N=1024 when the measured total was 22 GiB, so the ladder stopped early, and every reported 'capacity' was just a lower bound. Fix: project from the last measured total at the largest N, or from per-agent anonymous memory plus page cache counted once. Label skipped cells 'not run (projection)', never 'does not fit'. Also preserve expensive derived inputs (converted files) to storage before an auto-terminate-on-DONE watcher kills the VM.