diff --git a/CHANGELOG.md b/CHANGELOG.md
index 6790280..8695b25 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -21,6 +21,14 @@ All notable changes to styleprofile. The format follows
   applies eleven transforms (quotes, line endings, wrapping, dashes, ellipses, spacing,
   paragraph separators, an HTML round trip) and checks the verdicts. See "What styleprofile
   ignores" in `docs/method.md`.
+- **A flaky cache test no longer fails on busy machines.** `test_processes_can_share_the_cache`
+  sometimes found a writer's last entry missing on a loaded Windows runner. The cache did what
+  it promises: a writer stalled for over `HAND_DEADLINE_S` is given up on, its unwritten
+  entries are dropped and `problem` says so. The test asserted that nothing is ever dropped.
+  It now checks the real contract (no failure, a sound file, correct entries, every entry
+  present unless a writer reported skipping, the file usable afterwards), its processes use
+  generous timeouts, and two new tests cover a lock released in time and one held too long.
+  The cache itself is unchanged.
 
 ## [0.2.0] - 2026-10-08
 
diff --git a/tests/test_cache.py b/tests/test_cache.py
index a72a8b8..cdeccf2 100644
--- a/tests/test_cache.py
+++ b/tests/test_cache.py
@@ -309,30 +309,131 @@ def value(number: int) -> dict[str, Any]:
     assert path.stat().st_size < 1_000_000
 
 
+# The cache may skip writes by design: a writer that makes no progress for HAND_DEADLINE_S (5 s,
+# less than the 10 s TIMEOUT_S it may legitimately wait for another process's lock) is given up
+# on and what is unwritten dropped, with ``problem`` saying so. A busy CI runner (Windows jobs
+# under ``pytest -n auto``) can stall a writer that long. These processes therefore raise the
+# limits far past anything a loaded runner needs, so that skipping is not what this test sees,
+# and report ``problem`` so that the test still checks the contract if one is skipped anyway.
 WRITER_SCRIPT = """
 import sys
 from pathlib import Path
-from styleprofile.cache import MeasurementCache
-store = MeasurementCache(Path(sys.argv[1]))
+from styleprofile import cache
+cache.TIMEOUT_S = cache.HAND_DEADLINE_S = 120.0
+cache.CLOSE_TIMEOUT_S = 180.0
+store = cache.MeasurementCache(Path(sys.argv[1]))
 for number in range(300):
     store.put(f"{sys.argv[2]}{number}".encode(), number, {"n": number})
 store.close()
+print(store.problem or "")
 """
 
 
 def test_processes_can_share_the_cache(tmp_path: Path) -> None:
+    """Three processes write to one file. The contract: no process fails, the file stays sound,
+    what is in it is right, and every entry is there unless that process reported (``problem``)
+    that the cache skipped its writes. Afterwards the file can be read and written again."""
     path = tmp_path / "cache.sqlite3"
+    names = ("a", "b", "c")
     processes = [
-        subprocess.Popen([sys.executable, "-c", WRITER_SCRIPT, str(path), name])
-        for name in ("a", "b", "c")
+        subprocess.Popen(
+            [sys.executable, "-c", WRITER_SCRIPT, str(path), name],
+            stdout=subprocess.PIPE,
+            text=True,
+        )
+        for name in names
     ]
-    assert [process.wait(timeout=60) for process in processes] == [0, 0, 0]
+    problems = {}
+    for name, process in zip(names, processes, strict=True):
+        output, _ = process.communicate(timeout=240)
+        assert process.returncode == 0
+        problems[name] = output.strip()
+    with sqlite3.connect(path) as db:
+        assert db.execute("PRAGMA integrity_check").fetchall() == [("ok",)]
+    db.close()
     store = MeasurementCache(path)
-    for name in ("a", "b", "c"):
-        found = store.fetch(f"{name}299".encode())
-        assert found is not None and caching.decode(found[1]) == {"n": 299}
+    present = 0
+    for name in names:
+        for number in range(300):
+            found = store.fetch(f"{name}{number}".encode())
+            if found is not None:
+                assert caching.decode(found[1]) == {"n": number}
+                present += 1
+            else:
+                assert problems[name], f"{name}{number} is missing and {name} reported no problem"
+    assert store.problem is None
+    if not any(problems.values()):
+        assert present == 900 == caching.describe(path)[2]
+    store.put(b"after", 1, {"n": 1})
     store.close()
-    assert caching.describe(path)[2] == 900
+    store = MeasurementCache(path)
+    assert store.fetch(b"after") is not None
+    store.close()
+
+
+def _hold_lock(path: Path) -> sqlite3.Connection:
+    """Another process's write transaction, as far as the cache can tell."""
+    holder = sqlite3.connect(path, isolation_level=None, check_same_thread=False)
+    holder.execute("BEGIN EXCLUSIVE")
+    return holder
+
+
+def test_a_lock_released_in_time_loses_no_writes(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    path = tmp_path / "cache.sqlite3"
+    MeasurementCache(path).close()  # creates the file
+    monkeypatch.setattr(caching, "TIMEOUT_S", 60.0)
+    monkeypatch.setattr(caching, "HAND_DEADLINE_S", 60.0)
+    store = MeasurementCache(path)
+    holder = _hold_lock(path)
+    timer = threading.Timer(0.5, holder.execute, ["COMMIT"])
+    timer.start()
+    for number in range(300):
+        store.put(f"k{number}".encode(), number, {"n": number})
+    store.close()  # waits for the lock rather than giving up
+    timer.join()
+    holder.close()
+    assert store.problem is None
+    assert caching.describe(path)[2] == 300
+
+
+def test_a_lock_held_too_long_skips_the_writes_and_says_so(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """The writer may wait out the lock (TIMEOUT_S) longer than ``close`` waits for a writer
+    that finishes nothing (HAND_DEADLINE_S): then the run goes on without those writes, with
+    ``problem`` set, rather than stall. The file is untouched and works afterwards."""
+    path = tmp_path / "cache.sqlite3"
+    seed = MeasurementCache(path)
+    seed.put(b"seed", 1, {"n": 0})
+    seed.close()
+    monkeypatch.setattr(caching, "TIMEOUT_S", 60.0)
+    monkeypatch.setattr(caching, "HAND_DEADLINE_S", 0.5)
+    monkeypatch.setattr(caching, "HAND_TIMEOUT_S", 0.1)
+    store = MeasurementCache(path)
+    holder = _hold_lock(path)
+    try:
+        for number in range(300):
+            store.put(f"k{number}".encode(), number, {"n": number})
+        writer = store._writer
+        assert writer is not None  # started by the first full batch
+        started = time.monotonic()
+        store.close()  # gives up while the lock is still held
+        assert time.monotonic() - started < 30
+        assert store.problem is not None and "stopped responding" in store.problem
+    finally:
+        holder.execute("ROLLBACK")
+        holder.close()
+    writer.join(timeout=60)
+    after = MeasurementCache(path)
+    assert after.fetch(b"seed") is not None  # what was there is intact
+    after.put(b"after", 1, {"n": 1})
+    after.close()
+    assert after.problem is None
+    with sqlite3.connect(path) as db:
+        assert db.execute("PRAGMA integrity_check").fetchall() == [("ok",)]
+    db.close()
 
 
 def test_the_cache_command_shows_and_clears_it(