test: widen m9_2 cancel-race harvest margin (macOS CI flake)

The fake server sleeps 250ms before responding; a 350ms total budget
left 100ms for two pipe transits + thread scheduling, which flaked on
loaded macOS runners — and the queued-stdin-writer hop added by the
typing-perf arc narrows it further. The contract under test is
cancel-after-queue-before-manager-tick, which any
long-enough-for-the-response wait preserves.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

tests/m9_2_acceptance.rs

@@ -536,10 +536,15 @@ fn m9_2_cancelled_sibling_wins_over_queued_response() {
         .read_resource(sid, "file:///race")
         .expect("read c");
 
-    // Let the fake server finish its delayed response, then harvest
-    // the supervisor event queue without giving McpManager a chance
-    // to process it yet.
-    std::thread::sleep(Duration::from_millis(350));
+    // Let the fake server finish its delayed response (it sleeps
+    // 250ms), then harvest the supervisor event queue without giving
+    // McpManager a chance to process it yet. The margin over the
+    // fake's delay must absorb two pipe transits plus the request's
+    // queued-stdin-writer hop under CI load (a 100ms margin flaked on
+    // macOS runners); a generous wait does not weaken the contract —
+    // the race under test is cancel-AFTER-queue-BEFORE-manager-tick,
+    // which holds for any wait long enough for the response to land.
+    std::thread::sleep(Duration::from_millis(1000));
     sup.borrow_mut().tick();
 
     // Cancel only b after the response is queued but before