[wstein-agents]

Adds tests/cocalc/b6_adversary.py and runs it in CoCalc CI. It checks backup --restored-before TIME --parent ID adversarially.

Oracle: restoring the new snapshot must give exactly the tree that was on disk when the backup ran. Each scenario prepares a restored tree, backs it up with the option and again without it (the control), then compares both restores with the disk.

• In contract (must not diverge):
  • baseline;
  • writes shortly after TIME, and right after it (inside the 2 s margin);
  • an interrupted restore that is then rerun;
  • links and special files;
  • chmod passes, before and after TIME.
• Out of contract (reported, expected to diverge): a non-empty restore target holding stale files, the wrong --parent, and a killed restore that was not rerun.
• Hard checks:
  • a future --restored-before is rejected;
  • no control backup ever diverges;
  • no in-contract scenario diverges.
• Negative controls:
  • the baseline must actually reuse files, otherwise the option did nothing and the passes prove nothing;
  • stale_nonempty_target and wrong_parent must show a divergence, which proves the oracle can see stale reuse;
  • a binary without the option exits 2.

Local run against the published v0.11.4-cocalc.4 (x86_64 binary sha256 7d5c3020…4414): exit 0, no problems. In-contract scenarios reused 18–70 files with 0 divergent; the two out-of-contract cases each have 1 divergent file; the future TIME was rejected. With v0.11.4-cocalc.3 the script exits 2 ("binary has no --restored-before").

[wstein-agents]

Merged at the exact approved head 42848b9d60fcc1be9be3f8bb234f5d160d92955f (merge commit e5fa0aca18d2f179f3eec61aa71dd90cf225b449).

• Merge bar for this fork (William, 10-31): "i think it is fine to merge with @review approval because that is just codex."
• @review: APPROVE at exact head 42848b9d. That covers the full diff against cocalc/main 7396f0e2, and @review independently checked that the harness passed in all six jobs.
• CI: all required checks green at the head (run 41333856278: 24 pass, 1 skipped).
• Two documentation nits (the docstring's exit summary, and the wording in s_write_after_time_no_margin) follow in a small PR from @rustic-tests.

Owner: @ops-4 on alpha.cocalc.ai, project ops (c3337130)