diff --git a/backend/apps/nine_router/process.py b/backend/apps/nine_router/process.py
index 58b24e9e..ad006682 100644
--- a/backend/apps/nine_router/process.py
+++ b/backend/apps/nine_router/process.py
@@ -689,7 +689,10 @@ async def p_ensure_running_impl():
 
 def stop():
     """Stop the 9Router subprocess."""
-    global p_process, watchdog_task, p_death_watcher_task
+    global p_process, watchdog_task, p_death_watcher_task, p_is_running_last_ok
+    # Killing it is exactly what makes a cached "it's up" a lie, so drop the cache here rather than
+    # in each caller; a stale True made ensure_running() skip the restart and the bounce report success.
+    p_is_running_last_ok = 0.0
     # Cancel the healers FIRST or they would revive the router we're about to kill (shutdown = the one sanctioned "down").
     if watchdog_task is not None:
         watchdog_task.cancel()
diff --git a/backend/tests/test_router_stop_clears_liveness_cache.py b/backend/tests/test_router_stop_clears_liveness_cache.py
new file mode 100644
index 00000000..45cacb65
--- /dev/null
+++ b/backend/tests/test_router_stop_clears_liveness_cache.py
@@ -0,0 +1,83 @@
+"""A deliberate stop() must not leave is_running() answering True from its cache.
+
+The 10s positive cache in is_running() exists so a busy router isn't misread as dead. Three paths
+kill the router; death_watch and the dev stale-kill branch both drop the cache, but stop() (the
+DELIBERATE kill) did not, and it cancels the death watcher first so nothing could cover for it.
+
+Live consequence, packaged 1.7.10-exp.2, 2026-08-31: bounce_router_after_connect ran stop() ->
+ensure_running(), the cached True made ensure_running log "9Router already running" and start
+nothing, and the bounce then returned is_running() == True. The router was down 22s, every CLI that
+connected in that window deferred its MCP tools, and the first ToolSearch load 400'd the turn
+(ENG-394) after 61 tool calls of real work.
+"""
+import time
+
+import pytest
+
+from backend.apps.nine_router import process as rp
+
+
+@pytest.fixture(autouse=True)
+def p_restore():
+    before = (rp.p_is_running_last_ok, rp.p_process, rp.watchdog_task, rp.p_death_watcher_task)
+    yield
+    (rp.p_is_running_last_ok, rp.p_process, rp.watchdog_task, rp.p_death_watcher_task) = before
+
+
+def p_arm_cache_and_kill_port(monkeypatch):
+    """Cache says 'up'; the port is dead. Only the cache can produce a True now."""
+    def p_refuse(*a, **k):
+        raise OSError("connection refused")
+    monkeypatch.setattr(rp.socket, "create_connection", p_refuse)
+    rp.p_is_running_last_ok = time.monotonic()
+    rp.p_process = None
+    rp.watchdog_task = None
+    rp.p_death_watcher_task = None
+
+
+def test_the_cache_is_what_makes_this_test_meaningful(monkeypatch):
+    """Control: with the cache armed and the port dead, is_running() DOES lie. If this ever fails
+    the cache is gone and the rest of the file proves nothing."""
+    p_arm_cache_and_kill_port(monkeypatch)
+    assert rp.is_running() is True
+
+
+def test_stop_clears_the_cache_so_a_dead_router_reads_as_dead(monkeypatch):
+    p_arm_cache_and_kill_port(monkeypatch)
+    rp.stop()
+    assert rp.is_running() is False, "stop() left is_running() answering True on a dead port"
+
+
+def test_stop_clears_the_cache_before_it_touches_anything_else():
+    """Ordering, not just behaviour: an edit that moves the reset below an early return, or into one
+    of the branches, fails here rather than shipping."""
+    import inspect
+    src = inspect.getsource(rp.stop)
+    reset = src.index("p_is_running_last_ok = 0.0")
+    first_branch = src.index("if watchdog_task is not None")
+    assert reset < first_branch, "the cache reset must sit above every branch in stop()"
+
+
+@pytest.mark.asyncio
+async def test_a_bounce_cannot_report_success_after_killing_the_router(monkeypatch):
+    """The row-6 half: the bounce used to return True having started nothing."""
+    from backend.apps.nine_router import bounce_after_connect as bac
+
+    p_arm_cache_and_kill_port(monkeypatch)
+    started: list[bool] = []
+
+    async def p_ensure():
+        # A real ensure_running only starts when is_running() is False; mirror that decision here.
+        if not rp.is_running():
+            started.append(True)
+
+    monkeypatch.setattr(rp, "ensure_running", p_ensure)
+    monkeypatch.setattr(bac.router_process, "ensure_running", p_ensure)
+    monkeypatch.setattr(bac.asyncio, "sleep", lambda *_a, **_k: p_noop())
+
+    async def p_noop():
+        return None
+
+    ok = await bac.bounce_router_after_connect("claude")
+    assert started == [True], "the bounce skipped the restart because the cache still said 'up'"
+    assert ok is False, "the bounce claimed success while the router was down"