diff --git a/tests/KNOWN_FAILURES.md b/tests/KNOWN_FAILURES.md
new file mode 100644
index 000000000..414c82059
--- /dev/null
+++ b/tests/KNOWN_FAILURES.md
@@ -0,0 +1,106 @@
+# Known full-suite failures
+
+`python -m pytest -q` does not come back clean on every machine, and it never
+has. Without a list of which failures are expected, a first local run is
+uninterpretable: you cannot tell "you broke something" from "you are on a Mac",
+so the usual result is either chasing a non-bug or ignoring a real one.
+
+This is that list. It is a record of observation, not a permission slip: a test
+here is still a test that does not pass, and three of the six below are
+defects someone should fix.
+
+Last measured: `lab @ c499c01b`, macOS 15 on Apple Silicon, Python 3.11.
+
+```
+6 failed, 10658 passed, 6 skipped
+```
+
+## Get the prerequisites right first
+
+Most "surprise" failures are a missing dependency rather than anything in this
+file. A clean run needs all of:
+
+```bash
+python3.11 -m venv venv
+./venv/bin/python -m pip install -r requirements.txt
+npm ci                              # the browser tests shell out to node
+npx playwright install chromium     # ~30 tests drive a real browser
+mkdir -p data                       # SQLite lives at ./data/app.db
+```
+
+plus `ffmpeg` on `PATH` for the media tests.
+
+If you already have a ChromaDB running, point `CHROMADB_PORT` at a closed port
+for the run. The client reaches Chroma over HTTP regardless of the data
+directory, so a test run will otherwise attach to whatever store is listening,
+including one holding real data.
+
+Miss `npm ci` and roughly 36 browser tests fail on `Cannot find package
+'playwright'`. That is not a regression, it is the missing install.
+
+## The six
+
+### Test bugs: comparing an unresolved path against a resolved one
+
+- `tests/test_code_nav_tools.py::test_read_file_extracts_structured_documents`
+- `tests/test_code_nav_tools.py::test_read_file_extracts_legacy_word_documents`
+- `tests/test_workspace_confine.py::test_glob_confined_e2e`
+
+```
+assert [('/private/tmp/codenav_.../report.docx', ...)]
+    == [('/tmp/codenav_.../report.docx', ...)]
+```
+
+On macOS `/tmp` is a symlink to `/private/tmp`. The code under test resolves
+the path and the assertion does not, so the two disagree about a file they both
+found. Nothing is wrong with the behaviour.
+
+**These are fixable and should be fixed**: resolve both sides before comparing.
+They are listed as known rather than environmental because the platform is only
+what exposes them.
+
+### Optional dependency: ffmpeg without a WebP encoder
+
+- `tests/test_inspect_media_tool.py::test_inspect_media_exports_final_decodable_frame_at_exact_duration`
+
+```
+ffmpeg still extraction failed: Automatic encoder selection failed ...
+Error opening output files: Encoder not found
+```
+
+The test asks ffmpeg for a `.webp` still and asserts `exit_code == 0`. WebP
+encoding is a build option, and Homebrew's ffmpeg does not always carry it. CI
+installs a build that does, which is why this is green there.
+
+**Needs a decision**: skip when the encoder is absent, or fall back to PNG. The
+current shape asserts success from a codec that is not guaranteed present.
+
+### Environmental: real sockets
+
+- `tests/test_integration_api_call_ssrf.py::test_real_socket_falls_back_from_dead_first_to_live_second`
+
+```
+httpcore.ConnectTimeout / httpx.ConnectTimeout
+```
+
+Opens real sockets and depends on a connection to a dead address being refused
+quickly rather than hanging. Sandboxed and restricted-network machines time out
+instead. Genuinely environmental.
+
+### Unexplained: rich-text colour contrast
+
+- `tests/test_document_rich_color_reset_and_contrast.py::test_rich_colors_follow_theme_and_undo_as_one_edit`
+
+A Playwright run times out waiting for `#doc-email-richbody p` to contain a
+`span` after a colour is applied.
+
+**This one is not flaky.** Three consecutive runs failed identically, each at
+about 31 seconds. It was previously written off as timing noise and that was
+wrong. The cause is not established, and until it is, treat it as a possible
+real defect in the rich-text colour path rather than a platform artifact.
+
+## Keeping this current
+
+Re-measure on a clean checkout of `lab` with the prerequisites above, and
+update the header revision, the counts and any entry that changed. A failure
+that appears and is not listed here is a regression until shown otherwise.
diff --git a/tests/README.md b/tests/README.md
index 085cb5f84..9d5d1f80a 100644
--- a/tests/README.md
+++ b/tests/README.md
@@ -33,6 +33,8 @@ the sub-area. The `area_*` names are registered in `pyproject.toml`; the dynamic
 `sub_*` names are registered before collection by `pytest_configure` in
 `tests/conftest.py`, so unknown-mark warnings still flag genuine typos.
 
+The full suite does not come back clean on every machine. [KNOWN_FAILURES.md](KNOWN_FAILURES.md) lists which failures are expected, which are test bugs worth fixing, and the prerequisites a clean run needs; anything not on that list is a regression until shown otherwise.
+
 For common focused runs, use `tests/run_focus.py`. It validates area and
 sub-area names, accepts sub-areas with or without the `sub_` prefix, and passes
 extra pytest arguments after `--`: