diff --git a/.console/log.md b/.console/log.md
index d60bdfb..dc55562 100644
--- a/.console/log.md
+++ b/.console/log.md
@@ -2,6 +2,56 @@
 
 _Chronological continuity log. Decisions, stop points, what changed and why._
 
+## 2026-08-04 — fix(tests): stop the suite reading ambient env config
+
+`REPOGRAPH_BOUNDARY_ARTIFACT_FILE` leaked from the caller's shell into every test.
+Two that assert the "no artifact configured" path were directly contradicted by it:
+
+    test_reconcile.py::TestAC1SingleSourceOfTruth::test_no_artifact_no_scrub_targets
+    test_boundary_detectors.py::TestB2Required::test_b2_flags_missing_required_boundary_source
+
+The exposure was two modules, not the one reported — worth checking for, because
+the second was in a file that already knows about the variable (it calls
+`monkeypatch.setenv` at line 71) and still had a test that inherited it.
+
+Backwards in the way that matters: the variable is legitimately exported by anyone
+who runs the audit locally or pushes through `.hooks/pre-push`, and absent in CI. So
+a developer with a *working* setup saw red while CI stayed green. That is the
+failure mode that teaches people to ignore their own test results.
+
+Fixed with an autouse fixture in `tests/conftest.py` that clears the variable for
+every test, rather than a `delenv` at the two call sites. The defect is that the
+suite reads ambient config at all; patching the two known victims leaves the next
+artifact-sensitive test to rediscover it. Tests that *want* the variable still set it
+explicitly with `monkeypatch.setenv`, which is unaffected.
+
+New `tests/test_env_isolation.py` pins the fixture, because without it a later
+refactor could drop the fixture and the only symptom would be a suite that passes in
+CI and fails on the machines of the people most likely to run it. `conftest` builds
+its list from `boundary._ARTIFACT_FILE_ENV` rather than a literal, so renaming the
+variable in the detector cannot leave the isolation silently covering nothing.
+
+Two self-inflicted detours worth recording, both caught by tooling rather than by
+me:
+
+- The first draft asserted only on `os.environ` and tripped our own **T8** — a test
+  file importing nothing from any src package. Fair catch; it was testing Python, not
+  Custodian.
+- The second imported the list via `from tests.conftest import ...` and passed
+  locally but **failed collection in CI**: `No module named 'tests'`. `tests/` has no
+  `__init__.py`, so that name resolves only when the repo root is on `sys.path` —
+  true under `python -m pytest`, which is what I verified with, and false under the
+  bare `pytest` CI runs. The lesson is the same one this entry is about: verifying
+  through a different entry point than CI uses hides exactly the class of bug being
+  fixed. Resolved by inverting the dependency — conftest derives the name from the
+  detector, and the test imports only from `custodian`, so no test→conftest import
+  exists to be fragile.
+
+Verified through the CI invocation this time — bare `pytest -q`, with the variable
+set and unset, plus `python -m pytest`: 1241 passed, 5 skipped, identical across all
+three. Pre-existing at origin/main; not caused by #72, which observed it and
+deliberately left it alone to stay focused.
+
 ## 2026-08-04 — chore(config): raise our own r1_line_budget to 1000, and say why
 
 `.console/log.md` sat at 396 against a 400 budget, so the next entry anyone wrote
diff --git a/tests/conftest.py b/tests/conftest.py
index d98b98d..e9d6035 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -6,6 +6,10 @@
 import sys
 from pathlib import Path
 
+import pytest
+
+from custodian.audit_kit.detectors.boundary import _ARTIFACT_FILE_ENV
+
 _REPO_ROOT = Path(__file__).parent.parent.resolve()
 _EXPECTED_VENV = (_REPO_ROOT / ".venv").resolve()
 _ACTIVE_PREFIX = Path(sys.prefix).resolve()
@@ -21,3 +25,30 @@
         f"Or invoke pytest through the venv directly:\n"
         f"  .venv/bin/pytest"
     )
+
+
+# Ambient environment that must never reach a test. Anyone who runs the audit
+# locally, or pushes through .hooks/pre-push, legitimately has this exported —
+# so a developer with a WORKING setup saw a red suite while CI stayed green,
+# which is the wrong way round and trains people to ignore failures.
+#
+# Two tests asserted behaviour for the "no artifact configured" case and were
+# contradicted by the inherited value:
+#   test_reconcile.py::TestAC1SingleSourceOfTruth::test_no_artifact_no_scrub_targets
+#   test_boundary_detectors.py::TestB2Required::test_b2_flags_missing_required_boundary_source
+#
+# Cleared for every test rather than patched at those two call sites: the bug is
+# that the suite reads ambient config at all, and a per-test fix leaves the next
+# artifact-sensitive test to rediscover it. Tests that WANT the variable set it
+# explicitly with monkeypatch.setenv (see test_boundary_detectors.py), which still
+# works — this only removes what leaked in from the caller's shell.
+#
+# Taken from the detector rather than spelled as a literal: boundary.py owns the
+# name, so renaming it there cannot leave this isolation silently covering nothing.
+_AMBIENT_ENV_VARS = (_ARTIFACT_FILE_ENV,)
+
+
+@pytest.fixture(autouse=True)
+def _isolate_ambient_env(monkeypatch: pytest.MonkeyPatch) -> None:
+    for name in _AMBIENT_ENV_VARS:
+        monkeypatch.delenv(name, raising=False)
diff --git a/tests/test_env_isolation.py b/tests/test_env_isolation.py
new file mode 100644
index 0000000..3137c69
--- /dev/null
+++ b/tests/test_env_isolation.py
@@ -0,0 +1,53 @@
+# SPDX-License-Identifier: AGPL-3.0-or-later
+# Copyright (C) 2026 ProtocolWarden
+"""The suite must not read ambient environment config.
+
+Anyone who runs the audit locally, or pushes through ``.hooks/pre-push``,
+legitimately has ``REPOGRAPH_BOUNDARY_ARTIFACT_FILE`` exported. Before the autouse
+fixture in ``conftest.py`` that leaked into every test, so two tests asserting the
+"no artifact configured" case failed for a developer with a WORKING setup while CI
+— which has no such variable — stayed green. That is the wrong way round: it trains
+people to ignore failures.
+
+These tests pin the fixture. Without them a later refactor could drop it and the
+only symptom would be a suite that passes in CI and fails on the machines of the
+people most likely to run it.
+
+Imports the variable name from the detector that owns it, never from ``conftest``:
+``tests/`` is not a package, so ``from tests.conftest import ...`` resolves only when
+the repo root happens to be on ``sys.path`` — true under ``python -m pytest``, false
+under the bare ``pytest`` CI runs.
+"""
+from __future__ import annotations
+
+import os
+
+import pytest
+
+from custodian.audit_kit.detectors.boundary import _ARTIFACT_FILE_ENV
+
+
+def test_ambient_var_is_cleared_for_every_test():
+    """The autouse fixture applies here without this test requesting it."""
+    assert _ARTIFACT_FILE_ENV not in os.environ
+
+
+def test_the_cleared_name_is_the_one_the_detector_reads():
+    """Guard against the isolation drifting off the real variable.
+
+    ``conftest`` builds its list from this same constant, so this asserts the
+    wiring rather than a duplicated string.
+    """
+    assert _ARTIFACT_FILE_ENV == "REPOGRAPH_BOUNDARY_ARTIFACT_FILE"
+
+
+def test_a_test_can_still_opt_in_explicitly():
+    """Clearing ambient config must not stop a test setting the var on purpose.
+
+    ``test_boundary_detectors.py`` does exactly this to exercise the
+    artifact-configured path; the fixture must not fight it.
+    """
+    with pytest.MonkeyPatch.context() as mp:
+        mp.setenv(_ARTIFACT_FILE_ENV, "/some/explicit/path.json")
+        assert os.environ[_ARTIFACT_FILE_ENV] == "/some/explicit/path.json"
+    assert _ARTIFACT_FILE_ENV not in os.environ