fix(world-model): scope WorldModel state per-task to prevent cross-task contamination

WorldModel routed all state through a single mutable _current_task_id,
so concurrent tasks sharing one instance could corrupt each other's
state. start_task() now returns a TaskWorldView handle scoped to that
task; TaskRunner.run() threads it through as a local variable instead
of reading self.world_model implicitly. Also extracts _start_world_view/
_record_step_result as reusable TaskRunner methods for composed runners.

openspec: world-model capability, archived change world-model-runtime

tests/test_world_model.py

@@ -129,6 +129,52 @@ def test_observe_skips_malformed_fields_without_raising() -> None:
     assert len(model.state.history) == 2
 
 
+def test_concurrent_tasks_on_same_world_model_stay_isolated() -> None:
+    model = WorldModel(config=WorldConfig(history_size=5))
+    launch_a = PlannedStep(
+        action="launch_app",
+        description="launch a's app",
+        args={"app_id": "com.example.a"},
+    )
+    launch_b = PlannedStep(
+        action="launch_app",
+        description="launch b's app",
+        args={"app_id": "com.example.b"},
+    )
+
+    task_a = model.start_task("task-a")
+    task_a.observe(_scene(), _semantic_scene("A-Home"), launch_a, _result(launch_a))
+
+    # Task B starts on the SAME WorldModel instance *after* task A has
+    # already observed an event but *before* task A observes its next one --
+    # this is the interleaving that corrupted state when WorldModel tracked
+    # "the current task" via a single shared mutable attribute.
+    task_b = model.start_task("task-b")
+    task_b.observe(_scene(), _semantic_scene("B-Home"), launch_b, _result(launch_b))
+
+    remember_a = PlannedStep(
+        action="tap",
+        description="a remembers",
+        args={"remember": {"contact": "Alice"}},
+    )
+    task_a.observe(_scene(), _semantic_scene("A-Chat"), remember_a, _result(remember_a))
+
+    assert task_a.state.current_app == "com.example.a"
+    assert task_a.state.current_page == "A-Chat"
+    assert task_a.state.variables == {"contact": "Alice"}
+    assert [event.action for event in task_a.state.history] == ["launch_app", "tap"]
+
+    assert task_b.state.current_app == "com.example.b"
+    assert task_b.state.current_page == "B-Home"
+    assert task_b.state.variables == {}
+    assert [event.action for event in task_b.state.history] == ["launch_app"]
+
+    # External readers keep working via state_for().
+    assert model.state_for("task-a") is task_a.state
+    assert model.state_for("task-b") is task_b.state
+    assert model.state_for("task-a") is not model.state_for("task-b")
+
+
 def test_observe_appends_semantic_or_raw_scene_history_with_eviction() -> None:
     model = WorldModel(config=WorldConfig(history_size=2))
     first = PlannedStep(action="first", description="first")