Fix transient NO_SWITCHER / CDP race in box-chat-cdp.py

Three fixes for flaky main-chat reads (pip/opm):
1. ENSURE: retry chat-switcher lookup 4x with 2s waits instead of
   immediate NO_SWITCHER (React may still be rendering).
2. ev(): drain CDP events until matching command id arrives;
   previously the first recv() could grab a browser event instead
   of our evaluate response, returning None.
3. main(): reconnect fresh websocket on each retry (3 attempts);
   reusing a stale ws after page navigation gave dead JS contexts.

Verified: 7/8 reads succeed across all 4 agents; the 1 failure
was THREAD_NOT_FOUND while opm was actively in a side chat,
self-healed on next attempt.

Session: sidechat/chromebox-ops

bin/box-chat-cdp.py

@@ -55,7 +55,22 @@ def ev(ws, expr, await_p=False):
         "params": {"expression": expr, "returnByValue": True,
                    "awaitPromise": await_p},
     }))
-    resp = json.loads(ws.recv())
+    # Drain CDP events until we get our command response (id 1).
+    # The browser can emit events (Runtime.executionContextCreated, etc.)
+    # at any time; taking the first recv() blindly returns None on a
+    # busy page (observed as transient NO_SWITCHER / "unexpected
+    # messages payload" on pip/opm 2026-10-04).
+    resp = None
+    drained = 0
+    for _ in range(50):
+        raw = ws.recv()
+        resp = json.loads(raw)
+        if resp.get("id") == 1:
+            break
+        drained += 1
+    else:
+        raise RuntimeError("CDP: no response to Runtime.evaluate (drained %d)" % drained)
+    pass  # drained count available in `drained` if needed
     res = resp.get("result", {})
     if res.get("subtype") == "error":
         raise RuntimeError("JS error: %s" % str(res.get("description"))[:200])
@@ -79,7 +94,13 @@ ENSURE = """(async () => {
   if (nav.getAttribute('aria-current') !== 'page') { nav.click(); await sleep(3000); }
   const panelOpen = () => !!document.querySelector('[data-testid="hatch-chat-compose"]');
   if (!panelOpen()) {
-    const sw = document.querySelector('[data-testid="hatch-chat-switcher-trigger"]');
+    // Retry: the switcher may not be rendered yet if the SPA is still
+    // settling (observed transient NO_SWITCHER on pip/opm 2026-10-04).
+    let sw = null;
+    for (let k = 0; k < 4 && !sw; k++) {
+      sw = document.querySelector('[data-testid="hatch-chat-switcher-trigger"]');
+      if (!sw) await sleep(2000);
+    }
     if (!sw) return 'NO_SWITCHER';
     sw.click(); await sleep(2500);
     if (!panelOpen()) return 'PANEL_CLOSED';
@@ -250,12 +271,19 @@ def main(argv):
     agent, op = argv[1], argv[2]
     ws = None
     try:
-        ws = connect(agent)
-        # CDP evaluate can rarely resolve null when the page is mid-navigation;
-        # one retry covers the flake (observed 2026-10-04).
+        # CDP evaluate can resolve null when the page is mid-navigation
+        # (agent actively using browser). Reconnect fresh on each retry
+        # so we attach to the current page, not a stale JS context.
+        # Observed 2026-10-04: opm's browser navigates during reads.
         data, last_err = None, None
-        for attempt in range(2):
+        for attempt in range(3):
             try:
+                if ws is not None:
+                    try:
+                        ws.close()
+                    except Exception:
+                        pass
+                ws = connect(agent)
                 if op == "threads":
                     data = op_threads(ws)
                     ok = isinstance(data, dict) and isinstance(