test(infra): poll-for-event race fixes + watchdog timeout bump + spec update

tests/conftest.py

@@ -181,7 +181,7 @@ def _check_required_test_dependencies() -> None:
   raise pytest.UsageError(msg)
 
 def _smart_watchdog_exit() -> None:
- if not _pytest_finished_event.wait(timeout=600.0):
+ if not _pytest_finished_event.wait(timeout=900.0):
   os._exit(2)
  import time
  time.sleep(5.0)

tests/test_reset_session_clears_mma_and_rag.py

@@ -26,10 +26,27 @@ def test_reset_session_clears_mma_tier_usage(live_gui) -> None:
   'tier_usage': {'Tier 1': {'model': 'polluted'}},
   'tickets': []
  })
- time.sleep(0.5)
+ # Poll until the polluted entry is visible. Without this, the reset
+ # can fire BEFORE the push_event task is processed (async via io_pool
+ # + GUI render loop), and the test would falsely pass even if the
+ # reset didn't actually clear anything.
+ for _ in range(40):
+  state = client.get_gui_state()
+  mma = state.get('mma_state', {})
+  tier1 = mma.get('tier_usage', {}).get('Tier 1', {})
+  if tier1.get('model') == 'polluted':
+   break
+  time.sleep(0.25)
  # Trigger the reset
  client.reset_session()
- time.sleep(0.5)
+ # Poll until the polluted entry is gone
+ for _ in range(40):
+  state = client.get_gui_state()
+  mma = state.get('mma_state', {})
+  tier1 = mma.get('tier_usage', {}).get('Tier 1', {})
+  if tier1.get('model') != 'polluted':
+   break
+  time.sleep(0.25)
  # Verify the polluted entry is gone
  state = client.get_gui_state()
  mma = state.get('mma_state', {})
@@ -50,10 +67,21 @@ def test_reset_session_clears_mma_status(live_gui) -> None:
   'tier_usage': {},
   'tickets': []
  })
- time.sleep(0.5)
+ # Poll for the polluted status to be visible BEFORE the reset
+ for _ in range(40):
+  state = client.get_gui_state()
+  if state.get('mma_status') == 'running':
+   break
+  time.sleep(0.25)
  client.reset_session()
- time.sleep(0.5)
- state = client.get_gui_state()
+ # Poll for the reset to have taken effect. Without this, the
+ # mma_state_update task can fire AFTER the reset, setting status
+ # back to 'running' (race condition surfaced in batched live_gui).
+ for _ in range(40):
+  state = client.get_gui_state()
+  if state.get('mma_status') == 'idle':
+   break
+  time.sleep(0.25)
  assert state.get('mma_status') == 'idle', (
   f"mma_status not reset: {state.get('mma_status')!r}"
  )
@@ -70,10 +98,19 @@ def test_reset_session_clears_active_tier(live_gui) -> None:
   'tier_usage': {},
   'tickets': []
  })
- time.sleep(0.5)
+ # Poll for the polluted active_tier to be visible BEFORE the reset
+ for _ in range(40):
+  state = client.get_gui_state()
+  if state.get('active_tier') == 'Tier 2 (Tech Lead)':
+   break
+  time.sleep(0.25)
  client.reset_session()
- time.sleep(0.5)
- state = client.get_gui_state()
+ # Poll for the reset to have taken effect
+ for _ in range(40):
+  state = client.get_gui_state()
+  if state.get('active_tier') is None:
+   break
+  time.sleep(0.25)
  assert state.get('active_tier') is None, (
   f"active_tier not reset: {state.get('active_tier')!r}"
  )

tests/test_z_negative_flows.py

@@ -112,10 +112,13 @@ def test_mock_timeout(live_gui) -> None:
   client.set_value("ai_input", "Trigger timeout")
   client.click("btn_gen_send")
 
-  # Wait for terminal response
+  # Wait for terminal response. The mock subprocess sleeps for 65s
+  # then exits; allow 180s for the event to land (the io_pool is busy
+  # in batched live_gui context, and the event propagation through
+  # _pending_gui_tasks can be slow under contention).
   event = None
   start = time.time()
-  while time.time() - start < 80:
+  while time.time() - start < 180:
    ev = client.wait_for_event("response", timeout=5)
    if ev and ev.get("payload", {}).get("status") != "streaming...":
     event = ev