test(init): resolve sys.path entries against cwd so the suite is green from a worktree (#454)

FORK_CHANGELOG.md

@@ -56,6 +56,48 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
   *Files:* `docs/fork-changes/`, `scripts/fork_changes.py`, `scripts/render-docs.py`, `scripts/maintain-fork-changes.py`, `scripts/check-docs.sh`
 
 
+### Fixed
+
+
+- **test_init's sys.path assertion resolves entries against cwd, so it stops failing in every linked worktree** ([`HEAD`](https://github.com/techempower-org/mempalace/commit/HEAD))
+  `test_init_filters_sys_path_from_leaked_pythonpath` failed all five sentinel
+  params in any linked worktree, so every lane in the 2026-09-10 drain wave ran
+  the suite with it deselected and paid a false red before working that out.
+
+  The behaviour under test was correct throughout; the assertion was not. It
+  iterated `for p in sys.path if p`, excluding the empty string — and the empty
+  string is the cwd marker `python -c` puts on the path, which is what actually
+  resolves the child's import. Measured in both trees with PYTHONPATH set to
+  the sentinel, as the test sets it: the child imports the tree it was launched
+  from either way (`EMPTY_IN_PATH: True` in both), but `PARENT_ON_SYS_PATH` was
+  True in the main tree and False from a worktree. The absolute-entry match
+  succeeded in the main tree for an incidental reason — the shared venv's
+  editable install adds exactly that directory to `sys.path` — so the green was
+  measuring "is this the editable target?", not "did the filter over-strip?".
+  From a worktree the editable entry points at the main tree instead, the
+  absolute match fails, and the test reports an over-strip that never happened.
+
+  Entries are now normalised with `os.path.abspath(p or os.curdir)` before
+  comparison, which is faithful to the assertion's stated purpose — the package
+  must stay importable, so its parent must survive on `sys.path` — because the
+  cwd entry is how it stayed importable. It keeps its teeth: the assertion can
+  only fail if the filter removed both the cwd marker and any absolute entry
+  providing the package, which is exactly an over-strip, and a child that
+  cannot import at all still trips the `returncode == 0` assertion first.
+
+  `MEMPALACE_IMPORTED_FROM` is printed for diagnosis and deliberately not
+  asserted on: which tree the child resolves depends on its cwd, and pinning
+  that would re-introduce the coupling being removed.
+
+  Verified both directions, since a fix that only passes where it already
+  passed would be worthless: 8/8 from a linked worktree and 8/8 with cwd set to
+  the main tree, against 5 failed / 3 passed for the old test from a worktree.
+  The full suite now runs green from a worktree with no manual deselect.
+
+  *Tests:* 0 new (repairs 5 existing params; the suite from a worktree goes 5 failed -> 0)
+  *Files:* `tests/test_init.py`
+
+
 ## [2026-09-10]
 
 

README.md

@@ -290,6 +290,7 @@ The full enumeration of fork-ahead changes. The canonical source is [`docs/fork-
 
 | Description | Upstream PR | Fork commit |
 |---|---|---|
+| test_init's sys.path assertion resolves entries against cwd, so it stops failing in every linked worktree | — | [`HEAD`](https://github.com/techempower-org/mempalace/commit/HEAD) |
 | Fork-change entries split one-per-file; entry shas verified by ancestry and resolved after merge | — | [`HEAD`](https://github.com/techempower-org/mempalace/commit/HEAD) |
 | README install path documents putting the mempalace-mcp bridge on PATH | — | [`933602e`](https://github.com/techempower-org/mempalace/commit/933602e) |
 | Checkpoint drawers carry the session id, not just the diary entry; session_id is validated | — | [`1a54838`](https://github.com/techempower-org/mempalace/commit/1a54838) |

docs/fork-changes/2026-09-11-init-test-cwd-relative-syspath.yaml

@@ -0,0 +1,46 @@
+seq: 139
+id: init-test-cwd-relative-syspath
+date: '2026-09-11'
+bucket: Fixed
+commit: HEAD
+fork_pr: 482
+area: Testing
+summary: test_init's sys.path assertion resolves entries against cwd, so it stops failing
+  in every linked worktree
+body: |
+  `test_init_filters_sys_path_from_leaked_pythonpath` failed all five sentinel
+  params in any linked worktree, so every lane in the 2026-09-10 drain wave ran
+  the suite with it deselected and paid a false red before working that out.
+
+  The behaviour under test was correct throughout; the assertion was not. It
+  iterated `for p in sys.path if p`, excluding the empty string — and the empty
+  string is the cwd marker `python -c` puts on the path, which is what actually
+  resolves the child's import. Measured in both trees with PYTHONPATH set to
+  the sentinel, as the test sets it: the child imports the tree it was launched
+  from either way (`EMPTY_IN_PATH: True` in both), but `PARENT_ON_SYS_PATH` was
+  True in the main tree and False from a worktree. The absolute-entry match
+  succeeded in the main tree for an incidental reason — the shared venv's
+  editable install adds exactly that directory to `sys.path` — so the green was
+  measuring "is this the editable target?", not "did the filter over-strip?".
+  From a worktree the editable entry points at the main tree instead, the
+  absolute match fails, and the test reports an over-strip that never happened.
+
+  Entries are now normalised with `os.path.abspath(p or os.curdir)` before
+  comparison, which is faithful to the assertion's stated purpose — the package
+  must stay importable, so its parent must survive on `sys.path` — because the
+  cwd entry is how it stayed importable. It keeps its teeth: the assertion can
+  only fail if the filter removed both the cwd marker and any absolute entry
+  providing the package, which is exactly an over-strip, and a child that
+  cannot import at all still trips the `returncode == 0` assertion first.
+
+  `MEMPALACE_IMPORTED_FROM` is printed for diagnosis and deliberately not
+  asserted on: which tree the child resolves depends on its cwd, and pinning
+  that would re-introduce the coupling being removed.
+
+  Verified both directions, since a fix that only passes where it already
+  passed would be worthless: 8/8 from a linked worktree and 8/8 with cwd set to
+  the main tree, against 5 failed / 3 passed for the old test from a worktree.
+  The full suite now runs green from a worktree with no manual deselect.
+tests: 0 new (repairs 5 existing params; the suite from a worktree goes 5 failed -> 0)
+files:
+- tests/test_init.py

tests/test_init.py

@@ -34,7 +34,20 @@ def test_init_filters_sys_path_from_leaked_pythonpath(pythonpath):
     Asserts on the sentinel substring directly so the test does not
     couple to the production normalization logic. The dot/empty/unset
     cases additionally exercise the early-return / collision paths
-    without crashing."""
+    without crashing.
+
+    The parent-retained assertion resolves relative ``sys.path`` entries
+    against the child's cwd before comparing (#454). ``python -c`` puts the
+    cwd on ``sys.path`` as the empty string, and that entry is what actually
+    makes the package importable here — the child imports the tree it was
+    launched from. Comparing only ABSOLUTE entries made the assertion pass
+    for an incidental reason: in the main tree the editable install happens
+    to add that exact directory, so the check succeeded; from a linked
+    worktree the editable entry points at the main tree instead, the
+    absolute match failed, and all five sentinel params went red while the
+    behaviour under test was working perfectly. Measured both ways; every
+    lane on 2026-09-10 paid that false red.
+    """
     env = os.environ.copy()
     if pythonpath is None:
         env.pop("PYTHONPATH", None)
@@ -44,11 +57,12 @@ def test_init_filters_sys_path_from_leaked_pythonpath(pythonpath):
         "import mempalace, os, sys; "
         f"prefix = {_LEAK_PREFIX!r}; "
         "mempalace_parent = os.path.dirname(os.path.dirname(mempalace.__file__)); "
+        "norm = lambda p: os.path.normcase(os.path.normpath(os.path.abspath(p or os.curdir))); "
         "print('ENV:', repr(os.environ.get('PYTHONPATH'))); "
         "print('SENTINEL_IN_PATH:', any(prefix in (p or '') for p in sys.path)); "
         "print('MEMPALACE_PARENT_PRESENT:', any("
-        "os.path.normcase(os.path.normpath(p)) == os.path.normcase(os.path.normpath(mempalace_parent)) "
-        "for p in sys.path if p))"
+        "norm(p) == norm(mempalace_parent) for p in sys.path)); "
+        "print('MEMPALACE_IMPORTED_FROM:', mempalace_parent)"
     )
     result = subprocess.run(
         [sys.executable, "-c", code],
@@ -74,6 +88,10 @@ def test_init_filters_sys_path_from_leaked_pythonpath(pythonpath):
     assert "MEMPALACE_PARENT_PRESENT: True" in out, (
         f"filter over-stripped sys.path (mempalace parent gone): {diag}"
     )
+    # MEMPALACE_IMPORTED_FROM is printed for diagnosis and deliberately NOT
+    # asserted on: which tree the child resolves depends on its cwd, and
+    # pinning that is the very coupling #454 was about. It belongs in the
+    # failure output, not in the contract.
 
 
 def test_init_preserves_cwd_marker_when_pythonpath_collides():

website/public/llms-full.txt

@@ -309,6 +309,7 @@ The full enumeration of fork-ahead changes. The canonical source is [`docs/fork-
 
 | Description | Upstream PR | Fork commit |
 |---|---|---|
+| test_init's sys.path assertion resolves entries against cwd, so it stops failing in every linked worktree | — | [`HEAD`](https://github.com/techempower-org/mempalace/commit/HEAD) |
 | Fork-change entries split one-per-file; entry shas verified by ancestry and resolved after merge | — | [`HEAD`](https://github.com/techempower-org/mempalace/commit/HEAD) |
 | README install path documents putting the mempalace-mcp bridge on PATH | — | [`933602e`](https://github.com/techempower-org/mempalace/commit/933602e) |
 | Checkpoint drawers carry the session id, not just the diary entry; session_id is validated | — | [`1a54838`](https://github.com/techempower-org/mempalace/commit/1a54838) |