Skip to content

Make native Windows admission-fence tests independent of private task directories #108

Closed

Description

@CompleteDotTech

The native Windows admission-fence tests depend on a historical private execution directory instead of an owned disposable fixture. When that directory is absent, all nine cases fail in setUp() before exercising any fence behavior.

Fresh Windows Python 3.13 baseline at 04acabedcfff1cef9871087fd6a6374375ab5ebf: 1,404 tests, nine errors, eight skips. All nine errors point to the fixture directory creation in tests/test_paper_pilot_phase_cutover_guarded_v2.py:41.

The affected cases can be isolated with:

python -m unittest discover -s tests -p "test_paper_pilot_phase_cutover_guarded_v2.py" -v

Required change:

  • Use a unique owned temporary directory independent of historical task state.
  • Preserve the actual Windows ACL, contender, crash/replay, rename and unchanged-state assertions.
  • Preserve fixture cleanup without changing unrelated ACLs or runtime fence policy.
  • Run the nine cases on native Windows with supported interpreters and include a portable hosted Windows check.
  • Refresh delivery manifests, obtain Astra review of the final signed head, pass hosted checks, merge and verify afterward.

This is a test portability and coverage defect discovered during the authorized domain/verifier delivery. It grants no historical-pilot restart, namespace recreation or execution authority, and does not satisfy the empirical acceptance of #90/#23.

Activity

  1. added a commit that references this issue on Oct 8, 2026
  2. CompleteDotTech commented on Oct 8, 2026

    @CompleteDotTech
    CollaboratorAuthor

    Completed by PR #109, merge 143d6dc7805b74240e09b610b94ddbb893d68fd5.

    Acceptance verified:

    • Each test owns a unique disposable fixture, independent of historical task directories; protected ACL setup occurs before pinning and stays inside that fixture.
    • Original contender, ACL, crash/replay, rename, readable-state and unchanged-byte assertions remain. Strict production comparisons and cleanup boundaries were preserved.
    • Native Windows Python 3.12 and 3.13 each pass all 64 hosted cases without skips: original 9, trustee 14, fixture 13, strict 10 and diagnostics 18. Post-merge traces verify successful fencing and complete deny/directory cleanup.
    • Manifest/source archive audited: 709 entries / 710 members; SHA-256 34a733f411ada1c88419b5321d05f1a49a75b30333f932f09ebd0ff29f367a4c. Existing source-suffix exclusion is qualified.
    • Actual Sonnet 5.5 implementation; independent actual Astra (gpt-6-astra, session 01a1190e-8968-7163-b4fd-d073c1423d6c) APPROVE on exact signed head d5e101ff470b44cf851c1ffa485fb2dab3d7b65c; signed merge delivers the same tree. Attributed review: comment.
    • Merged-main CI run 37712250412 completed successfully: Linux 3.11/3.12/3.13 plus Windows 3.12/3.13. Post-merge local smoke passed 36 portable regressions with no source edits. Original workspace safely synchronized and clean.

    This closes the test-portability defect only. Historical pilot authority, private evidence and empirical acceptance of #90/#23 remain separate and open; no expired window was reopened.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions