CU-1: regression test, Cursor write from a HOME holding .chock is judged in the workspace

[jothimani-rajendran]

What

CU-1 root cause. On the owner's Windows machine, the installed chock-witness Cursor plugin was built by an engine older than 9914aa7 ("fix(gate): take Cursor's repository root from workspace_roots", 2026-10-07). The owner's diagnostic shows that build has no workspace_root. Cursor sends preToolUse without cwd, so that build's repo_root_for fell back to the hook process's cwd. When the hook ran from C:\Users\jothi, the write became witness-repo-cursor/.github/workflows/witness.yml, which is outside .github/workflows/*, so it was allowed. Main has had the fix since 9914aa7.

This PR adds the regression test for that exact layout, so it can't come back:

• HOME holds .chock/, and the repo sits beside it, with HEAD unborn or committed;
• the hook runs from HOME;
• the payload has no cwd, only workspace_roots (spelled /C:/… on Windows);
• an unpinned write must be denied, and clean content allowed.

It also adds CPython 3.14.3, the interpreter the owner's hooks launch, to the windows-paths matrix.

Fails before, passes after (measured locally, Linux)

• The engine at 9914aa7^ (pytest -o pythonpath="<9914aa7^ src> tools" … -k home_with_chock): AssertionError: assert 'allow' == 'deny', 2 failed.
• Main (a2dd3de): 2 passed.
• The Windows result is in this PR's windows-paths job.

Definition of done

• ruff check and ruff format clean on the touched test file
• actionlint clean; matrix change only; no new permissions or actions
• CI green, including windows-paths on 3.12.0 / 3.12 / 3.13 / 3.14.3 / 3.14
• Changelog fragment: none (test and CI only; the fix itself shipped in 9914aa7 with changelog.d/cursor-windows-workspace-root.md)

🤖 Generated with Claude Code

---

Generated by Claude Code