Research: Charter CLI tests independent of invoking checkout

R-01 — Where the fix belongs

  • Decision: Test-side isolation; no product change.
  • Rationale: In production the guard's probe point and repository-root resolution both derive from the process working directory, so they cannot disagree. They split only under the test patch seam re-exported at src/specify_cli/cli/commands/charter/__init__.py:16-23. find_repo_root follows a linked worktree to the repository root checkout (generate.py:431-437), so probing the resolved root would hide the condition the guard exists to catch.
  • Alternatives considered: (b) derive the probe point from the resolved root — rejected, reintroduces the split the guard closed. (c) stub resolve_charter_write_root in tests — rejected, removes the guard from the tests that exercise the write path.

R-02 — Shape of the shared fixture

  • Decision: Opt-in fixture returning a callable that sets working directory and find_repo_root to one project root; default tmp_path.
  • Rationale: The existing copies only chdir(tmp_path) and rely on each test's own patch; affected tests use different project roots (tmp_path / "mock_repo", plain tmp_path). A callable serves both. Opt-in because tests/cli/commands/test_charter_json_error_contract.py has tests that need the invoking checkout's working directory.
  • Alternatives considered: autouse isolation for all charter tests — rejected (C-003). A fourth per-file copy — rejected (C-004).

R-03 — Recurrence check: runtime tripwire over static census

  • Decision: Autouse call-through wrapper on the guard at its three call sites; violation when the probed path is inside the invoking checkout; failure at teardown.
  • Rationale: Post-spec review found five patch/invoke spellings across 21 files, mixed files, and helper-indirected invokes. A static per-file census is evadable; the runtime check is per test and spelling-independent, and it fires from a repository root checkout, where the leak is otherwise invisible. Precedent for an autouse checkout-related fixture in the root conftest: _neutralize_worktree_detection.
  • Alternatives considered: AST census with allowlist — rejected as evadable and as standing debt. Running the suite from a real worktree in CI — out of scope (cost).
  • Known limit: commands started as a separate process are not seen.

R-04 — Pinned-rule inventory

  • Finding: tests/release/pinning_rule_inventory.json lists tests/cli/commands/test_charter_json_error_contract.py::_isolate_cwd_from_worktree_guard (subject retiring-step, form docstring-or-comment, not a dependency) because the copy's docstring mentions the make test-fast target. tests/release/test_pinning_inventory_fresh.py requires the file to match its derivation.
  • Decision: Remove the copy, re-derive the inventory with scripts/ci/derive_pinning_inventory.py (no flag regenerates; --check verifies), and keep the make test-fast literal out of the new support modules so no new rule is minted. The script itself should need no edit.
  • Pre-existing red: the inventory is already stale on the base branch (a fresh derivation adds tests/_support/p0_repro.py::<module-docstring>); reported as #5660 and cleared by the same regeneration.

R-05 — Finding stragglers

  • Decision: One bounded discovery run in IC-03: the test files that reference the charter command app, started once with the tripwire on.
  • Rationale: The known set came from a linked-worktree run of those same files (about 4 minutes). The tripwire reports the same set from any checkout, so the run doubles as proof. It is a named-file run, not a directory or suite sweep.

R-06 — Smoke test remedy

  • Finding: The guard runs before the dry-run branch (synthesize.py:212 vs :393), so the subprocess must not start in a linked worktree. The test asserts Evidence dry-run summary and Code signals:.
  • Decision: Run the subprocess in a seeded tmp Python project; keep both assertions.

R-07 — Issues

  • Closes: #5317, #5601, and #5660 (the pinning inventory was already stale on the base branch; IC-01 regenerates it).
  • #5317's acceptance-test half: not reproducible on current main or at base 99352f08; verdict "verified, no work".
  • Not this cause: see #5140. Guard defects: see #5411, #4250. Earlier per-file fix: see #4873.