diff --git a/tests/conftest.py b/tests/conftest.py
index 6dd2e511..feb981e1 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -72,6 +72,50 @@ def _redirect_sessions_index(monkeypatch, tmp_path):
     )
 
 
+@pytest.fixture(autouse=True)
+def _redirect_the_config_path(monkeypatch, tmp_path):
+    """Point `config_path()` at this test's scratch tree, in every module that binds it.
+
+    The read half of what `_guard_real_config_files` enforces for writes. The
+    suite never intends to touch the host's `~/.emrg/config.toml`, and it was
+    reading it constantly: constructing a real `EmrgServer` resolves the default
+    path in `ConfigReloader.__init__` (it fingerprints the file as its baseline,
+    `fingerprint` reads bytes rather than statting). Measured with a spy on
+    `open`/`io.open` (2026-09-18): **202 reads of the host's real file per
+    full-suite run** — 201 attributed to that construction (`daemon.py:259` →
+    `config_reload.py:167` → the `fingerprint` read) and 1 to a poll that let the
+    reloader tick (`daemon.py:650`). Nothing failed, which is the point — a host
+    read that no assertion depends on is invisible until a later edit makes an
+    expectation depend on it, which is how `tests/test_ws_e2e.py`'s upgrade
+    isolation became dead code and how `tests/test_config_reload.py` inherited
+    14 host resolutions.
+
+    Counting `config_path()` calls is not the same instrument: `ConfigReloader`
+    resolves the path and then reads it, so a fixture that patched the reloader
+    would hide the read from that spy without ever stopping it. The default path
+    is only *used* when no explicit `path=` is passed — the upgrade loop's
+    `load_update_config()` is the other caller, and it is not reached by the
+    suite (that path's isolation is a separate red line).
+
+    Two modules are re-pointed, because neither name is the other's alias:
+    `emrg.config` (where `config_path` is defined, and where `load_config` /
+    `load_update_config` call it) and `emrg.server.config_reload`, which imported
+    the name by value — patching only the first leaves the reloader on the host.
+
+    The scratch path keeps the shape of the real one — `config.toml` under a
+    directory named `.emrg` — so the test whose subject *is* the default
+    resolution (`tests/test_config.py::test_config_path`) still asserts what it
+    means to. A test that wants a config file of its own patches `config_path`
+    after this fixture, as `tests/test_upgrade.py` already does.
+    """
+    import emrg.config as cfg_mod
+    import emrg.server.config_reload as cr_mod
+
+    cfg_path = tmp_path / ".emrg" / "config.toml"
+    monkeypatch.setattr(cfg_mod, "config_path", lambda: cfg_path)
+    monkeypatch.setattr(cr_mod, "config_path", lambda: cfg_path)
+
+
 @pytest.fixture(autouse=True)
 def _ensure_git_on_path(monkeypatch):
     """Make bare ``git`` subprocess calls work on hosts without PATH git.
diff --git a/tests/test_config.py b/tests/test_config.py
index e844294b..49917c8e 100644
--- a/tests/test_config.py
+++ b/tests/test_config.py
@@ -13,6 +13,8 @@
     load_config,
     ensure_config,
 )
+from emrg import config as cfg_mod
+from emrg.server import config_reload as cr
 
 
 def test_config_dir():
@@ -28,6 +30,45 @@ def test_config_path():
     assert p.parent.name == ".emrg"
 
 
+def test_the_default_config_path_is_redirected_into_the_scratch_tree(tmp_path):
+    """`conftest` redirects the default path; this pins that it is in force.
+
+    Read through the *module attribute*, not through this file's own
+    `from emrg.config import config_path` — a by-value import is a separate name
+    that a fixture patching `emrg.config` cannot reach, which is the trap this
+    whole change is about (and the reason this test failed the first time it was
+    written: it called its own copy and got the host's path back).
+
+    `test_config_path` above asserts the *shape* of the default resolution, and
+    the redirection keeps the shape (`config.toml` under a `.emrg` directory), so
+    it cannot tell a redirected run from an unredirected one. Without this pin the
+    fixture could be deleted and the only symptom would be the suite quietly
+    reading the host's real `~/.emrg/config.toml` again — 202 reads per run,
+    measured with a spy on `open`/`io.open` (2026-09-18) and 0 after the
+    redirection. Attribution of the 202: `daemon.py:259` (the `ConfigReloader`
+    the server builds) 201, `daemon.py:650` (a poll that ticks) 1.
+    """
+    p = cfg_mod.config_path()
+    assert p.parent.name == ".emrg"
+    assert p.parent.parent == tmp_path, (
+        "the suite resolved the default config path outside the test's scratch "
+        "tree — the host's real ~/.emrg/config.toml is being read again"
+    )
+
+
+def test_the_reloader_resolves_the_config_path_inside_the_scratch_tree(tmp_path):
+    """The same redirection for the other module that **binds** the name.
+
+    `config_reload.py` does `from emrg.config import ... config_path ...`, so it
+    holds its own reference; a fixture that re-pointed only `emrg.config` would
+    leave `ConfigReloader.__init__` fingerprinting the host's file. Pinned
+    separately because the two are separate names, and the difference is exactly
+    the class of defect this suite has hit twice (`tests/test_ws_e2e.py`'s
+    module-level binding, `tests/test_config_reload.py`'s 14 host reads).
+    """
+    assert cr.config_path() == tmp_path / ".emrg" / "config.toml"
+
+
 def test_llm_config_defaults():
     cfg = LlmConfig()
     assert cfg.base_url == "https://api.openai.com/v1"