Add 4 red tests for Phase 2: shared node pool + contextvar HUD

Tests that will pass once implemented:
- pool_creates_shared_nodes: NodePool has shared stateless nodes
- pool_excludes_stateful: sensor/memorizer/ui not shared
- pool_reuses_instances: same pool returns same objects
- contextvar_hud_isolation: concurrent tasks get isolated HUD

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

tests/test_engine.py

@@ -526,6 +526,76 @@ def test_model_override_per_request():
     assert result["trace"]["path"] == "reflex"
 
 
+# --- Phase 2: Shared Node Pool (RED — will fail until implemented) ---
+
+def test_pool_creates_shared_nodes():
+    """NodePool creates shared instances for stateless nodes."""
+    from agent.node_pool import NodePool
+    pool = NodePool("v4-eras")
+    # Shared nodes should exist
+    assert "input" in pool.shared, "input should be shared"
+    assert "output" in pool.shared, "output should be shared"
+    assert "pa" in pool.shared, "pa should be shared"
+    assert "expert_eras" in pool.shared, "expert_eras should be shared"
+    assert "interpreter" in pool.shared, "interpreter should be shared"
+
+
+def test_pool_excludes_stateful():
+    """NodePool excludes stateful nodes (sensor, memorizer, ui)."""
+    from agent.node_pool import NodePool
+    pool = NodePool("v4-eras")
+    assert "sensor" not in pool.shared, "sensor should NOT be shared"
+    assert "memorizer" not in pool.shared, "memorizer should NOT be shared"
+    assert "ui" not in pool.shared, "ui should NOT be shared"
+
+
+def test_pool_reuses_instances():
+    """Two Runtimes using the same pool share node objects."""
+    from agent.node_pool import NodePool
+    pool = NodePool("v4-eras")
+    # Same pool → same node instances
+    input1 = pool.shared["input"]
+    input2 = pool.shared["input"]
+    assert input1 is input2, "pool should return same instance"
+
+
+def test_contextvar_hud_isolation():
+    """Contextvars isolate HUD events between concurrent tasks."""
+    from agent.nodes.base import _current_hud
+
+    results_a = []
+    results_b = []
+
+    async def hud_a(data):
+        results_a.append(data)
+
+    async def hud_b(data):
+        results_b.append(data)
+
+    async def task_a():
+        _current_hud.set(hud_a)
+        # Simulate work with a yield point
+        await asyncio.sleep(0)
+        hud_fn = _current_hud.get()
+        await hud_fn({"from": "a"})
+
+    async def task_b():
+        _current_hud.set(hud_b)
+        await asyncio.sleep(0)
+        hud_fn = _current_hud.get()
+        await hud_fn({"from": "b"})
+
+    async def run_both():
+        await asyncio.gather(task_a(), task_b())
+
+    asyncio.get_event_loop().run_until_complete(run_both())
+
+    assert len(results_a) == 1 and results_a[0]["from"] == "a", \
+        f"task_a HUD leaked: {results_a}"
+    assert len(results_b) == 1 and results_b[0]["from"] == "b", \
+        f"task_b HUD leaked: {results_b}"
+
+
 # --- Test registry (for run_tests.py) ---
 
 TESTS = {
@@ -538,8 +608,13 @@ TESTS = {
     'frame_trace_reflex': test_frame_trace_reflex,
     'frame_trace_expert': test_frame_trace_expert,
     'frame_trace_expert_with_interpreter': test_frame_trace_expert_with_interpreter,
-    # Red — Phase 1: config-driven models
+    # Phase 1: config-driven models
     'graph_has_models': test_graph_has_models,
     'instantiate_applies_graph_models': test_instantiate_applies_graph_models,
     'model_override_per_request': test_model_override_per_request,
+    # Phase 2: shared node pool
+    'pool_creates_shared_nodes': test_pool_creates_shared_nodes,
+    'pool_excludes_stateful': test_pool_excludes_stateful,
+    'pool_reuses_instances': test_pool_reuses_instances,
+    'contextvar_hud_isolation': test_contextvar_hud_isolation,
 }