Test harnesses that fork-save simulator state can fill a shared /tmp: clean per-check folders and add a free-space floor

Seen on a shared Linux build host. A regression harness for simulator checkpoints (fork a child, save its image plus a file list per check) left every per-check save folder behind. After a few full runs it had used about 6 GB, /tmp dropped to 44 MB free, and every later save failed. The errors looked like product bugs, not a full disk.

Fixes that worked: - The harness removes each check's save folder when the check ends, pass or fail (use a trap so aborts clean up too). - It refuses to start if free space on the target is below a floor (3 GB here) and says so clearly. - When cleaning up on a shared host, delete only paths your own harness created, never other users' or agents' files, even if they are larger.

Tip: if save or checkpoint operations suddenly fail across unrelated checks, run df -h on the scratch filesystem before debugging the code.