test: keep the project-root test inside its fixture when TMPDIR is under $HOME

[RobertSigmundsson]

Second of the three small things from our 3.0.3 migration. Test-only, and it cost us a
puzzled half-hour before we spotted what was going on — which is usually a sign it is worth
fixing for whoever hits it next.

Summary

• One unit test fails on any machine whose TMPDIR sits under the home directory, because the
  code under test walks out of the fixture and into the real filesystem.
• Two lines of the fixture change; the assertion is untouched.
• Test-only. No production code changes.

Why

detect_project_root() walks up from Path.cwd() and stops only when it reaches Path.home()
(src/surreal_memory/surface/resolver.py:195-205):

for _ in range(20):
    if current != home and _is_project_root(current, global_dir):
        return current
    parent = current.parent
    if parent == current or current == home:
        break

test_relocated_global_config_dir_is_not_a_project_root patches cwd to tmp_path/srv/data and
home to tmp_path/home. That home is a sibling of the walk, not an ancestor of it, so the walk
never meets it. It climbs out of tmp_path, into the real home directory, finds a marker there
and returns it:

    assert detect_project_root() is None
E   AssertionError: assert PosixPath('<$HOME>') is None

The marker it finds is usually .surrealmemory — checked first in the marker list, and created
in the home directory by this tool itself — or package.json. Removing one does not help,
because the walk then matches the other.

pytest puts tmp_path under TMPDIR, so whether the test passes depends entirely on whether
the developer's TMPDIR happens to be outside their home directory. The sibling test at line 48
is unaffected because its cwd is the patched home, so it breaks out on the first iteration.

Changes

• tests/unit/test_surface_path_and_decay.py — anchor the patched home at tmp_path so the
  walk terminates inside the fixture, with a comment stating why. The assertion, the global-dir
  patching and what the test proves are all unchanged.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature
• Breaking change
• Documentation update
• Refactoring (no functional changes)
• Tests

Test plan

• pytest tests/ -m "not stress" -n auto passes locally — and now under both worlds:

  TMPDIR=/tmp             ->  6719 passed, 119 skipped, 1 xfailed
  TMPDIR under $HOME      ->  6719 passed, 119 skipped, 1 xfailed   (before: 1 failed)
  
  the single test file:
  TMPDIR under $HOME, before  ->  1 failed, 22 passed
  TMPDIR under $HOME, after   ->  23 passed
  TMPDIR=/tmp, after          ->  23 passed
  

• ruff check src/ tests/ clean on 0.16.1. On 0.15.20 the tree still reports the
  pre-existing S310 in mcp/version_check_handler.py, which this branch does not touch.

• mypy src/ --ignore-missing-imports clean — Success: no issues found in 351 source files.

• Verified in both directions, because a test that cannot fail proves nothing. With the
  global-config-dir guard in surface/resolver.py deliberately removed —

  ```python
  if marker == _NM_DIR and global_dir is not None:
      try:
          if marker_path.resolve() == global_dir:
              continue
  ```
  
  — the fixed test fails again under **both** `TMPDIR` settings (`1 failed, 1 passed`), and
  passes again once the guard is restored (`2 passed`). So the fix did not make it
  vacuously green; it still catches the regression it was written for.
  

• Manual check: applies cleanly onto a pristine v3.0.3 tree.

Checklist

• My code follows the project's style guidelines
• I have performed a self-review of my code
• I have added tests that prove my fix is effective — the change is to a test; its power is
  demonstrated by the guard-removal run above.
• New and existing unit tests pass locally with my changes
• I have updated the documentation accordingly — not applicable.
• I have updated CHANGELOG.md — deliberately not: test-only, no user-visible effect,
  matching #137 and #140. Say the word and I will add one.

Related Issues

None.

Verified by

@RobertSigmundsson