[Deep-Sixed]

Closes #38.

writable_scratch=False was ignored

The OCI backend always mounted /tmp, /home, /mnt, /root, /run and /var as writable tmpfs. With writable_scratch=False they are now mounted read-only (still empty), on top of the existing --read-only root filesystem, so /stele/output is the only writable path, as under bubblewrap with Landlock.

Exit statuses 125–127

These were always reported as engine_error, even when the parser itself exited with one of them. A status is now an engine_error only when stderr ends with the engine's or OCI runtime's own message:

• a failed exec of the command (OCI runtime ..., executable file not found in $PATH, unable to start container process, OCI permission denied);
• Error response from daemon;
• the Error: (Podman) or docker: prefix at 125, for unknown images or runtimes and the like.

Otherwise it is the parser's own exit_status. I captured Podman's real messages for each case locally.

Limit (documented in docs/containment.md): a parser that prints the engine's exact wording and exits 125 is still taken for an engine error. The alternative, a start marker written by an entrypoint wrapper, would change what runs inside every container, so I left it out.

Tests

• Unit: the argv for both scratch modes, and the classification of six engine messages and four parser exits.
• Live, on every OCI lane (runc, gVisor, Podman, Docker): a parser exiting 125, 126 and 127 is exit_status; with writable_scratch=False, writes to /tmp /var /run /home /root /mnt /etc / all fail and /stele/output works.
• Locally (Podman): the new live tests pass. The Docker and gVisor lanes run only in CI.
• Full suite: 879 passed, 97 skipped. ruff is clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_013p3FU6gUJ1MGinnmEgvHsm

---

Generated by Claude Code

[Deep-Sixed]

linux-containment failed on the first push, and it was a real bug in this PR. I fixed it in d18ae27.

The failure: test_container_engine_error on oci-runc-docker and oci-runsc-docker. A missing parser command (exit 127) came back as exit_status, not engine_error.

Why: I matched the engine's message against the last line of stderr. Podman's errors are a single line, which is all I could check locally. Docker ends its error with a second line, Run 'docker run --help' for more information, so the marker was never on the last line.

Fix: look at the last three non-empty lines. A marker far from the end (a parser's own noise) still doesn't count. The unit tests now include Docker's two-line form, plus a case with an engine-looking line earlier in the output. The full suite passes locally (882 passed, 97 skipped). The Docker lanes run only in CI, so I'm watching the new run.

---

Generated by Claude Code

---

Generated by Claude Code