feat(coord): Stage 3 SessionManager Children primitive lift + cluster bus push paths

Lift the Children primitive out of CoordinatorAdapter into universal
SessionManager core primitives, replace the fragile poll + state-event
piggyback paths with first-class cluster bus event types for inline
approval delivery, and clean up the resulting frontend reducer.

Architecture
- New `turnstone/core/children_registry.py` — universal parent → children
  + reverse-lookup primitive with atomic `add_child` (returns parent UI
  for race-free dispatch). Lifted from `CoordinatorAdapter`.
- New `turnstone/core/child_source.py` — `ChildSource` Protocol with
  `SameNodeChildSource` (in-process via SessionManager state observer)
  and `ClusterChildSource` (cross-node via ClusterCollector listener).
- `SessionManager._on_state_change` upgraded to multi-subscriber
  (`subscribe_to_state` / `unsubscribe_from_state`) under a dedicated
  lock; CLI consumer migrated.
- `CoordinatorAdapter` shrunk: 731 → ~640 LOC. Children data lives in
  the registry; fan-out lives in ClusterChildSource. Backward-compat
  property facades dropped; tests updated to use the registry surface.

Cluster bus event vocabulary
- New event types `intent_verdict`, `approval_resolved`,
  `approve_request` flow through both `ClusterCollector._apply_delta`
  (translation from node SSE) and `emit_console_ws_*` (synthesis on
  console pseudo-node).
- `CoordinatorAdapter._dispatch_child_event` re-emits as
  `child_ws_intent_verdict` / `child_ws_approval_resolved` /
  `child_ws_approve_request` on the parent coord's SSE stream.
- New `_broadcast_intent_verdict` / `_broadcast_approval_resolved` /
  `_broadcast_approve_request` no-op hooks on `SessionUIBase`. WebUI
  pushes to the global queue; ConsoleCoordinatorUI pushes to the
  collector. `approve_tools` calls `_broadcast_approve_request` right
  after setting `_pending_approval` so the items reach the coord tree
  immediately, eliminating the bulk-fetch race.

Cleanups
- `pending_approval_detail` piggyback on `ws_state` / `cluster_state`
  removed end-to-end. Bulk fetch + explicit verdict / approve-request
  push are the canonical carriers.
- Browser `_judgePollTick` 90-second poll loop deleted; push path is
  authoritative.
- `urgent` flag on `scheduleLiveFetch` deleted (only caller was 409
  retry; replaced with `invalidateLiveBadge` + standard schedule).
- Console `_fetch_live_block` derives `pending_approval` from a
  disjunction (`activity_state="approval"` OR `state="attention"`
  OR detail present) so the bulk fetch can't return false during the
  state-transition race window.
- Coord-side merge guard in `flushLiveFetches` no longer clobbered:
  `handleChildState` only stamps `sseUpdatedAt` when authoritatively
  clearing detail.
- `child_locality` capability flag removed (was inert dead code).

Reliability
- Selective drop on listener queue overflow: critical event types
  (verdicts, approvals, ws_closed, child_ws_*) evict one oldest item
  to make room rather than dropping themselves on a full queue.
  Best-effort events (state ticks, content tokens, status, activity)
  drop as before. Applied to `SessionUIBase._enqueue`,
  `ClusterCollector._fanout`, and the `WebUI._global_queue` puts in
  the new broadcast hooks.
- `_state_subscribers` snapshot under a dedicated lock so concurrent
  subscribe / unsubscribe during dispatch can't shift the iterator.

UX / a11y
- Loading placeholder in renderChildRow keeps row height stable while
  the bulk fetch is in-flight (sr-friendly aria-label).
- Focus preservation across `_renderChildrenNow` (capture +
  restore by row + marker) and across targeted `_updateChildRow` swaps.
- Layout-shift transition on the approval block max-height; respects
  `prefers-reduced-motion`.
- Sidebar pending count: `(N children · M pending)`.
- Risk pill `aria-label` spells out level + confidence for SR users.
- Per-coord SSE listener queue depth surfaced in the status bar
  (`queue N/500`) with color escalation (warn at >50%, danger at >80%).

Tests
- 305+ test changes across 8 files. New unit tests for
  `ChildrenRegistry`, `ChildSource` (both impls + multi-subscriber
  observer), the new collector emit + apply_delta cases, the dispatch
  cases for new event types, the broadcast hook overrides on both
  WebUI and ConsoleCoordinatorUI, and the focus / placeholder /
  pending-count frontend assertions in `test_coordinator_page.py`.

5024 passed, ruff + mypy clean.

tests/_coord_test_helpers.py

@@ -35,11 +35,10 @@ def _seed_children(
     The production path populates the registry via the cluster-event
     fan-out thread observing ``ws_created`` events. These tests just
     need a known-children set for the endpoint handlers to iterate —
-    inject directly under ``_children_lock`` rather than spinning up
-    the collector + fan-out plumbing.
+    inject directly via the registry's bulk-merge surface rather than
+    spinning up the collector + fan-out plumbing.
     """
-    with adapter._children_lock:
-        adapter._merge_child_ids_locked(coord_ws_id, child_ws_ids)
+    adapter._registry.merge_children(coord_ws_id, child_ws_ids)
 
 
 class _AuthMiddleware(BaseHTTPMiddleware):

tests/test_child_source.py

@@ -0,0 +1,270 @@
+"""Unit tests for :mod:`turnstone.core.child_source`.
+
+Covers both strategies in isolation against fakes — no live collector,
+no live SessionManager. Adapter-level integration coverage continues to
+live in ``test_coordinator_adapter.py``.
+"""
+
+from __future__ import annotations
+
+import contextlib
+import time
+from typing import TYPE_CHECKING, Any
+
+from turnstone.core.child_source import ClusterChildSource, SameNodeChildSource
+from turnstone.core.children_registry import ChildrenRegistry
+from turnstone.core.workstream import WorkstreamState
+
+if TYPE_CHECKING:
+    import queue
+
+
+# ---------------------------------------------------------------------------
+# SameNodeChildSource
+# ---------------------------------------------------------------------------
+
+
+class _FakeManager:
+    """Minimal SessionManager stand-in implementing the subscribe API."""
+
+    def __init__(self) -> None:
+        self.subscribers: list[Any] = []
+
+    def subscribe_to_state(self, callback: Any) -> None:
+        self.subscribers.append(callback)
+
+    def unsubscribe_from_state(self, callback: Any) -> None:
+        with contextlib.suppress(ValueError):
+            self.subscribers.remove(callback)
+
+    def fire(self, ws_id: str, state: WorkstreamState) -> None:
+        for cb in self.subscribers:
+            cb(ws_id, state)
+
+
+class TestSameNodeChildSource:
+    def test_start_subscribes_to_manager(self) -> None:
+        mgr = _FakeManager()
+        registry = ChildrenRegistry()
+        src = SameNodeChildSource(mgr, registry)
+        sink_calls: list[dict[str, Any]] = []
+        src.start(sink=sink_calls.append)
+        assert len(mgr.subscribers) == 1
+
+    def test_state_change_for_known_child_pushes_to_sink(self) -> None:
+        mgr = _FakeManager()
+        registry = ChildrenRegistry()
+        registry.install("p1", object())
+        registry.add_child("p1", "c1")
+        src = SameNodeChildSource(mgr, registry)
+        sink_calls: list[dict[str, Any]] = []
+        src.start(sink=sink_calls.append)
+
+        mgr.fire("c1", WorkstreamState.RUNNING)
+
+        assert len(sink_calls) == 1
+        ev = sink_calls[0]
+        assert ev["type"] == "cluster_state"
+        assert ev["ws_id"] == "c1"
+        assert ev["state"] == "running"
+
+    def test_state_change_for_unknown_workstream_is_dropped(self) -> None:
+        mgr = _FakeManager()
+        registry = ChildrenRegistry()
+        src = SameNodeChildSource(mgr, registry)
+        sink_calls: list[dict[str, Any]] = []
+        src.start(sink=sink_calls.append)
+
+        # No registry entry — pre-filter drops the event without
+        # invoking the sink.
+        mgr.fire("ws-unknown", WorkstreamState.IDLE)
+        assert sink_calls == []
+
+    def test_shutdown_unsubscribes(self) -> None:
+        mgr = _FakeManager()
+        registry = ChildrenRegistry()
+        src = SameNodeChildSource(mgr, registry)
+        src.start(sink=lambda ev: None)
+        assert len(mgr.subscribers) == 1
+        src.shutdown()
+        assert mgr.subscribers == []
+
+    def test_start_is_idempotent(self) -> None:
+        mgr = _FakeManager()
+        registry = ChildrenRegistry()
+        src = SameNodeChildSource(mgr, registry)
+        src.start(sink=lambda ev: None)
+        src.start(sink=lambda ev: None)
+        # Second start is a no-op; only one subscription.
+        assert len(mgr.subscribers) == 1
+
+    def test_sink_exception_does_not_propagate(self) -> None:
+        mgr = _FakeManager()
+        registry = ChildrenRegistry()
+        registry.install("p1", object())
+        registry.add_child("p1", "c1")
+        src = SameNodeChildSource(mgr, registry)
+
+        def bad_sink(ev: dict[str, Any]) -> None:
+            raise RuntimeError("sink boom")
+
+        src.start(sink=bad_sink)
+        # Should not raise — the strategy catches sink failures and logs.
+        mgr.fire("c1", WorkstreamState.RUNNING)
+
+
+# ---------------------------------------------------------------------------
+# ClusterChildSource
+# ---------------------------------------------------------------------------
+
+
+class _FakeCollector:
+    """Minimal ClusterCollector stand-in providing the listener API."""
+
+    def __init__(self, snapshot: dict[str, Any] | None = None) -> None:
+        self._snapshot = snapshot or {"nodes": []}
+        self.queues: list[queue.Queue[dict[str, Any]]] = []
+        self.unregistered: list[queue.Queue[dict[str, Any]]] = []
+
+    def get_snapshot_and_register(self, q: queue.Queue[dict[str, Any]]) -> dict[str, Any]:
+        self.queues.append(q)
+        return self._snapshot
+
+    def unregister_listener(self, q: queue.Queue[dict[str, Any]]) -> None:
+        self.unregistered.append(q)
+
+    def emit(self, event: dict[str, Any]) -> None:
+        """Push an event to all registered listener queues."""
+        for q in self.queues:
+            q.put(event)
+
+
+class TestClusterChildSource:
+    def test_start_subscribes_to_collector(self) -> None:
+        coll = _FakeCollector()
+        registry = ChildrenRegistry()
+        src = ClusterChildSource(
+            collector=coll,
+            registry=registry,
+            parents_provider=list,
+        )
+        try:
+            src.start(sink=lambda ev: None)
+            assert len(coll.queues) == 1
+        finally:
+            src.shutdown()
+
+    def test_start_primes_registry_from_snapshot(self) -> None:
+        snapshot = {
+            "nodes": [
+                {
+                    "workstreams": [
+                        {"id": "c1", "parent_ws_id": "p1"},
+                        {"id": "c2", "parent_ws_id": "p1"},
+                        # Unknown parent — dropped
+                        {"id": "x", "parent_ws_id": "p-unknown"},
+                    ],
+                },
+            ],
+        }
+        coll = _FakeCollector(snapshot)
+        registry = ChildrenRegistry()
+        registry.install("p1", object())
+        src = ClusterChildSource(
+            collector=coll,
+            registry=registry,
+            parents_provider=lambda: ["p1"],
+        )
+        try:
+            src.start(sink=lambda ev: None)
+            assert set(registry.children_of("p1")) == {"c1", "c2"}
+            assert registry.parent_for("x") is None
+        finally:
+            src.shutdown()
+
+    def test_event_dispatched_to_sink(self) -> None:
+        coll = _FakeCollector()
+        registry = ChildrenRegistry()
+        src = ClusterChildSource(
+            collector=coll,
+            registry=registry,
+            parents_provider=list,
+        )
+        sink_calls: list[dict[str, Any]] = []
+
+        try:
+            src.start(sink=sink_calls.append)
+            coll.emit({"type": "cluster_state", "ws_id": "c1", "state": "running"})
+            # Daemon thread loop has 1.0s queue timeout; poll briefly.
+            for _ in range(20):
+                if sink_calls:
+                    break
+                time.sleep(0.05)
+            assert len(sink_calls) == 1
+            assert sink_calls[0]["ws_id"] == "c1"
+        finally:
+            src.shutdown()
+
+    def test_shutdown_unregisters_and_joins_thread(self) -> None:
+        coll = _FakeCollector()
+        registry = ChildrenRegistry()
+        src = ClusterChildSource(
+            collector=coll,
+            registry=registry,
+            parents_provider=list,
+        )
+        src.start(sink=lambda ev: None)
+        src.shutdown()
+        assert coll.unregistered == coll.queues
+        # Second shutdown is a no-op (idempotent).
+        src.shutdown()
+
+    def test_start_is_idempotent(self) -> None:
+        coll = _FakeCollector()
+        registry = ChildrenRegistry()
+        src = ClusterChildSource(
+            collector=coll,
+            registry=registry,
+            parents_provider=list,
+        )
+        try:
+            src.start(sink=lambda ev: None)
+            src.start(sink=lambda ev: None)
+            assert len(coll.queues) == 1
+        finally:
+            src.shutdown()
+
+    def test_sink_exception_does_not_kill_thread(self) -> None:
+        coll = _FakeCollector()
+        registry = ChildrenRegistry()
+        src = ClusterChildSource(
+            collector=coll,
+            registry=registry,
+            parents_provider=list,
+        )
+        survived_calls: list[dict[str, Any]] = []
+        call_count = [0]
+
+        def flaky_sink(ev: dict[str, Any]) -> None:
+            call_count[0] += 1
+            if call_count[0] == 1:
+                raise RuntimeError("first one boom")
+            survived_calls.append(ev)
+
+        try:
+            src.start(sink=flaky_sink)
+            coll.emit({"type": "cluster_state", "ws_id": "c1", "state": "x"})
+            coll.emit({"type": "cluster_state", "ws_id": "c2", "state": "y"})
+            for _ in range(40):
+                if survived_calls:
+                    break
+                time.sleep(0.05)
+            assert len(survived_calls) == 1
+            assert survived_calls[0]["ws_id"] == "c2"
+        finally:
+            src.shutdown()
+
+
+# Multi-subscriber observer tests for ``SessionManager.subscribe_to_state``
+# / ``unsubscribe_from_state`` live in ``test_session_manager.py`` where
+# the proper FakeAdapter / FakeStorage construction helpers already exist.

tests/test_children_registry.py

@@ -0,0 +1,239 @@
+"""Unit tests for :class:`turnstone.core.children_registry.ChildrenRegistry`.
+
+The registry was lifted from ``CoordinatorAdapter`` in Stage 3 Step 1.
+Adapter-level coverage for the integrated behavior already lives in
+``test_coordinator_adapter.py``; this file pins the data structure
+invariants in isolation so the registry can be reused by future
+``ChildSource`` strategies (Step 2) without re-deriving the behavior
+from the adapter test surface.
+"""
+
+from __future__ import annotations
+
+import threading
+
+import pytest
+
+from turnstone.core.children_registry import ChildrenRegistry
+
+
+class _Sentinel:
+    """Lightweight UI stand-in; identity-comparable, no behavior."""
+
+
+@pytest.fixture
+def registry() -> ChildrenRegistry:
+    return ChildrenRegistry()
+
+
+# ---------------------------------------------------------------------------
+# install / uninstall
+# ---------------------------------------------------------------------------
+
+
+class TestInstallUninstall:
+    def test_install_seeds_empty_child_set_and_presence(self, registry: ChildrenRegistry) -> None:
+        ui = _Sentinel()
+        registry.install("p1", ui)
+        assert registry.children_of("p1") == []
+        assert registry.ui_for("p1") is ui
+        assert registry.parents() == ["p1"]
+
+    def test_install_is_idempotent_repoints_ui_keeps_children(
+        self, registry: ChildrenRegistry
+    ) -> None:
+        ui_a = _Sentinel()
+        ui_b = _Sentinel()
+        registry.install("p1", ui_a)
+        registry.merge_children("p1", ["c1", "c2"])
+        registry.install("p1", ui_b)
+        assert registry.ui_for("p1") is ui_b
+        assert set(registry.children_of("p1")) == {"c1", "c2"}
+
+    def test_uninstall_clears_forward_reverse_and_presence(
+        self, registry: ChildrenRegistry
+    ) -> None:
+        ui = _Sentinel()
+        registry.install("p1", ui)
+        registry.merge_children("p1", ["c1", "c2"])
+        registry.uninstall("p1")
+        assert registry.children_of("p1") == []
+        assert registry.ui_for("p1") is None
+        assert registry.parents() == []
+        assert registry.parent_for("c1") is None
+        assert registry.parent_for("c2") is None
+
+    def test_uninstall_unknown_parent_is_noop(self, registry: ChildrenRegistry) -> None:
+        registry.uninstall("never-installed")  # must not raise
+
+    def test_uninstall_does_not_clobber_other_parents(self, registry: ChildrenRegistry) -> None:
+        registry.install("p1", _Sentinel())
+        registry.install("p2", _Sentinel())
+        registry.merge_children("p1", ["c1"])
+        registry.merge_children("p2", ["c2"])
+        registry.uninstall("p1")
+        assert registry.parent_for("c1") is None
+        assert registry.parent_for("c2") == "p2"
+        assert registry.parents() == ["p2"]
+
+
+# ---------------------------------------------------------------------------
+# add_child — atomic check-and-route
+# ---------------------------------------------------------------------------
+
+
+class TestAddChild:
+    def test_add_child_returns_ui_on_success(self, registry: ChildrenRegistry) -> None:
+        ui = _Sentinel()
+        registry.install("p1", ui)
+        assert registry.add_child("p1", "c1") is ui
+        assert registry.parent_for("c1") == "p1"
+        assert registry.children_of("p1") == ["c1"]
+
+    def test_add_child_returns_none_when_parent_not_installed(
+        self, registry: ChildrenRegistry
+    ) -> None:
+        assert registry.add_child("absent", "c1") is None
+        assert registry.parent_for("c1") is None
+
+    def test_add_child_returns_none_on_duplicate(self, registry: ChildrenRegistry) -> None:
+        ui = _Sentinel()
+        registry.install("p1", ui)
+        assert registry.add_child("p1", "c1") is ui
+        # second add for same child returns None — caller must not
+        # double-dispatch.
+        assert registry.add_child("p1", "c1") is None
+        assert registry.children_of("p1") == ["c1"]
+
+
+# ---------------------------------------------------------------------------
+# merge_children — bulk seeding
+# ---------------------------------------------------------------------------
+
+
+class TestMergeChildren:
+    def test_merge_seeds_forward_and_reverse(self, registry: ChildrenRegistry) -> None:
+        registry.merge_children("p1", ["c1", "c2", "c3"])
+        assert set(registry.children_of("p1")) == {"c1", "c2", "c3"}
+        for cid in ("c1", "c2", "c3"):
+            assert registry.parent_for(cid) == "p1"
+
+    def test_merge_is_idempotent(self, registry: ChildrenRegistry) -> None:
+        registry.merge_children("p1", ["c1"])
+        registry.merge_children("p1", ["c1"])
+        assert registry.children_of("p1") == ["c1"]
+
+    def test_merge_skips_empty_or_falsy_ids(self, registry: ChildrenRegistry) -> None:
+        registry.merge_children("p1", ["", "c1", "", "c2"])
+        assert set(registry.children_of("p1")) == {"c1", "c2"}
+
+    def test_merge_does_not_require_install(self, registry: ChildrenRegistry) -> None:
+        # Snapshot-priming may run before the parent's install fires —
+        # the merge still seeds the forward set so the install picks
+        # the children up. (Storage-seeded rebuild relies on this.)
+        registry.merge_children("p1", ["c1"])
+        assert registry.children_of("p1") == ["c1"]
+        # ui_for is still None because install hasn't run
+        assert registry.ui_for("p1") is None
+
+
+# ---------------------------------------------------------------------------
+# Lookups — return copies, not live refs
+# ---------------------------------------------------------------------------
+
+
+class TestLookups:
+    def test_children_of_returns_copy(self, registry: ChildrenRegistry) -> None:
+        registry.install("p1", _Sentinel())
+        registry.merge_children("p1", ["c1", "c2"])
+        snap = registry.children_of("p1")
+        snap.append("c3-injected")
+        assert "c3-injected" not in registry.children_of("p1")
+
+    def test_children_of_unknown_parent_returns_empty(self, registry: ChildrenRegistry) -> None:
+        assert registry.children_of("absent") == []
+
+    def test_parent_for_unknown_child_returns_none(self, registry: ChildrenRegistry) -> None:
+        assert registry.parent_for("absent") is None
+
+    def test_parents_returns_copy(self, registry: ChildrenRegistry) -> None:
+        registry.install("p1", _Sentinel())
+        snap = registry.parents()
+        snap.append("p2-injected")
+        assert "p2-injected" not in registry.parents()
+
+
+# ---------------------------------------------------------------------------
+# Concurrency — concurrent add_child must not exceed the unique-set
+# invariant or leave a half-installed reverse-index entry.
+# ---------------------------------------------------------------------------
+
+
+class TestConcurrency:
+    def test_concurrent_add_child_returns_ui_exactly_once_per_unique(
+        self, registry: ChildrenRegistry
+    ) -> None:
+        ui = _Sentinel()
+        registry.install("p1", ui)
+        results: list[object] = []
+        results_lock = threading.Lock()
+
+        def attempt_add(child_id: str) -> None:
+            r = registry.add_child("p1", child_id)
+            with results_lock:
+                results.append(r)
+
+        threads = [threading.Thread(target=attempt_add, args=("c1",)) for _ in range(20)]
+        for t in threads:
+            t.start()
+        for t in threads:
+            t.join()
+
+        # Exactly one thread sees the UI; the remaining 19 see None
+        # (duplicate). The forward + reverse indexes carry exactly one
+        # entry for c1.
+        successes = [r for r in results if r is ui]
+        nones = [r for r in results if r is None]
+        assert len(successes) == 1
+        assert len(nones) == 19
+        assert registry.children_of("p1") == ["c1"]
+        assert registry.parent_for("c1") == "p1"
+
+    def test_concurrent_install_and_add_child_no_resurrect(
+        self, registry: ChildrenRegistry
+    ) -> None:
+        # add_child racing with uninstall: either lands first (registry
+        # populated) or the parent is gone (returns None). Must NOT
+        # leave a forward-set entry without presence — that would be
+        # the "resurrected after close" leak the locked dispatch path
+        # was guarding against.
+        ui = _Sentinel()
+        registry.install("p1", ui)
+
+        outcomes: list[object] = []
+
+        def adder() -> None:
+            outcomes.append(registry.add_child("p1", "c1"))
+
+        def uninstaller() -> None:
+            registry.uninstall("p1")
+
+        threads = [
+            threading.Thread(target=adder),
+            threading.Thread(target=uninstaller),
+        ]
+        for t in threads:
+            t.start()
+        for t in threads:
+            t.join()
+
+        # If add_child landed first: c1 is in the forward set, then
+        # uninstall clears everything. End state: nothing.
+        # If uninstall landed first: add_child sees no presence,
+        # returns None, no entry added. End state: nothing.
+        # Either way, the leak invariant holds: child set is empty or
+        # parent is gone, never "child set populated but no presence".
+        children = registry.children_of("p1")
+        ui_present = registry.ui_for("p1") is not None
+        if children:
+            assert ui_present, "registry leaked: children set without presence"

tests/test_console.py

@@ -314,12 +314,14 @@ class TestCollectorSnapshot:
         assert event["ws_id"] == "ws1"
         assert event["state"] == "running"
 
-    def test_apply_snapshot_state_change_forwards_pending_approval_detail(self):
-        """Reconnect-via-snapshot is the resync path after every console
-        restart or network blip.  Without forwarding the field here,
-        a child sitting in approval-pending across the gap renders as
-        ``activity_state=approval`` with no buttons until the next
-        state change — broken UX during the most common re-sync event."""
+    def test_apply_snapshot_state_change_does_not_carry_pending_approval_detail(self):
+        """Stage 3 cleanup — the snapshot-resync cluster_state event no
+        longer piggybacks ``pending_approval_detail`` (the field is
+        gone from cluster_state entirely). On reconnect the browser's
+        bulk fetch — triggered by the ``activity_state="approval"``
+        transition in the reducer — pulls the items directly from
+        ``ui.serialize_pending_approval_detail()`` via the dashboard
+        endpoint."""
         c = _make_collector()
         c._nodes["node-a"] = NodeSnapshot(
             node_id="node-a",
@@ -329,10 +331,6 @@ class TestCollectorSnapshot:
         q: queue.Queue[dict] = queue.Queue()
         c.register_listener(q)
 
-        detail = {
-            "items": [{"call_id": "c1", "header": "tool x"}],
-            "judge_pending": False,
-        }
         c._apply_snapshot(
             "node-a",
             {
@@ -344,7 +342,6 @@ class TestCollectorSnapshot:
                         "name": "same",
                         "state": "running",
                         "activity_state": "approval",
-                        "pending_approval_detail": detail,
                     }
                 ],
                 "health": {},
@@ -354,7 +351,8 @@ class TestCollectorSnapshot:
 
         event = q.get_nowait()
         assert event["type"] == "cluster_state"
-        assert event["pending_approval_detail"] == detail
+        assert event["activity_state"] == "approval"
+        assert "pending_approval_detail" not in event
 
     def test_apply_snapshot_skips_empty_id_workstream(self):
         c = _make_collector()
@@ -401,12 +399,12 @@ class TestCollectorDelta:
         # Verify in-memory state was updated
         assert c._nodes["node-a"].workstreams["ws1"]["state"] == "running"
 
-    def test_apply_delta_ws_state_forwards_pending_approval_detail(self):
-        """The rich approval payload now travels on the cluster bus so
-        coord tabs can render inline approve/deny buttons in lockstep
-        with the activity_state transition.  Collector must forward
-        the field verbatim — the adapter does the child-routing on
-        top, but the bus carries the data."""
+    def test_apply_delta_ws_state_does_not_carry_pending_approval_detail(self):
+        """Stage 3 cleanup — ``cluster_state`` no longer carries the
+        ``pending_approval_detail`` piggyback. Approval items now arrive
+        via bulk fetch on activity_state transition; verdicts via the
+        explicit ``intent_verdict`` event class. Symmetric event flow,
+        no piggyback to dedupe against."""
         c = _make_collector()
         c._nodes["node-a"] = NodeSnapshot(
             node_id="node-a",
@@ -416,10 +414,6 @@ class TestCollectorDelta:
         q: queue.Queue[dict] = queue.Queue()
         c.register_listener(q)
 
-        detail = {
-            "items": [{"call_id": "c1", "header": "tool x"}],
-            "judge_pending": False,
-        }
         c._apply_delta(
             "node-a",
             {
@@ -427,13 +421,13 @@ class TestCollectorDelta:
                 "ws_id": "ws1",
                 "state": "running",
                 "activity_state": "approval",
-                "pending_approval_detail": detail,
             },
         )
 
         event = q.get_nowait()
         assert event["type"] == "cluster_state"
-        assert event["pending_approval_detail"] == detail
+        assert event["activity_state"] == "approval"
+        assert "pending_approval_detail" not in event
 
     def test_apply_delta_ws_created(self):
         c = _make_collector()
@@ -480,6 +474,139 @@ class TestCollectorDelta:
         assert event["name"] == "new-name"
         assert c._nodes["node-a"].workstreams["ws1"]["name"] == "new-name"
 
+    def test_apply_delta_intent_verdict_forwards_verbatim(self):
+        """Stage 3 Step 5 — node-emitted intent_verdict events flow
+        through _apply_delta to cluster fan-out so coord adapters can
+        re-emit as child_ws_intent_verdict on the parent's SSE."""
+        c = _make_collector()
+        c._nodes["node-a"] = NodeSnapshot(
+            node_id="node-a",
+            server_url="http://a:8080",
+            workstreams={"ws1": {"id": "ws1", "name": "test", "state": "idle"}},
+        )
+        q: queue.Queue[dict] = queue.Queue()
+        c.register_listener(q)
+
+        verdict = {
+            "call_id": "c1",
+            "risk_level": "low",
+            "confidence": 0.9,
+            "recommendation": "approve",
+        }
+        c._apply_delta(
+            "node-a",
+            {"type": "intent_verdict", "ws_id": "ws1", "verdict": verdict},
+        )
+
+        event = q.get_nowait()
+        assert event["type"] == "intent_verdict"
+        assert event["ws_id"] == "ws1"
+        assert event["node_id"] == "node-a"
+        assert event["verdict"] == verdict
+
+    def test_apply_delta_intent_verdict_drops_when_ws_id_missing(self):
+        c = _make_collector()
+        c._nodes["node-a"] = NodeSnapshot(node_id="node-a", server_url="http://a:8080")
+        q: queue.Queue[dict] = queue.Queue()
+        c.register_listener(q)
+
+        c._apply_delta("node-a", {"type": "intent_verdict", "verdict": {}})
+
+        assert q.empty()
+
+    def test_apply_delta_approval_resolved_forwards_verbatim(self):
+        """Stage 3 Step 5 — paired with intent_verdict; clears the
+        coord tree's pending-approval pill in lockstep with the
+        actual decision rather than waiting for the state-change
+        piggyback."""
+        c = _make_collector()
+        c._nodes["node-a"] = NodeSnapshot(
+            node_id="node-a",
+            server_url="http://a:8080",
+            workstreams={"ws1": {"id": "ws1", "name": "test", "state": "idle"}},
+        )
+        q: queue.Queue[dict] = queue.Queue()
+        c.register_listener(q)
+
+        c._apply_delta(
+            "node-a",
+            {
+                "type": "approval_resolved",
+                "ws_id": "ws1",
+                "approved": True,
+                "feedback": "lgtm",
+                "always": False,
+            },
+        )
+
+        event = q.get_nowait()
+        assert event["type"] == "approval_resolved"
+        assert event["ws_id"] == "ws1"
+        assert event["node_id"] == "node-a"
+        assert event["approved"] is True
+        assert event["feedback"] == "lgtm"
+        assert event["always"] is False
+
+    def test_apply_delta_approve_request_forwards_detail(self):
+        """Push path for the initial approval items — eliminates the
+        bulk-fetch race that left the coord row stuck on a loading
+        placeholder when the bulk fetch landed in the gap between
+        _emit_state(ATTENTION) and approve_tools setting _pending_approval."""
+        c = _make_collector()
+        c._nodes["node-a"] = NodeSnapshot(
+            node_id="node-a",
+            server_url="http://a:8080",
+            workstreams={"ws1": {"id": "ws1", "name": "test", "state": "idle"}},
+        )
+        q: queue.Queue[dict] = queue.Queue()
+        c.register_listener(q)
+
+        detail = {
+            "type": "approve_request",
+            "items": [{"call_id": "c1", "header": "tool x"}],
+            "judge_pending": True,
+        }
+        c._apply_delta(
+            "node-a",
+            {"type": "approve_request", "ws_id": "ws1", "detail": detail},
+        )
+
+        event = q.get_nowait()
+        assert event["type"] == "approve_request"
+        assert event["ws_id"] == "ws1"
+        assert event["node_id"] == "node-a"
+        assert event["detail"] == detail
+
+    def test_apply_delta_approve_request_drops_when_ws_id_missing(self):
+        c = _make_collector()
+        c._nodes["node-a"] = NodeSnapshot(node_id="node-a", server_url="http://a:8080")
+        q: queue.Queue[dict] = queue.Queue()
+        c.register_listener(q)
+
+        c._apply_delta("node-a", {"type": "approve_request", "detail": {}})
+
+        assert q.empty()
+
+    def test_apply_delta_approval_resolved_coerces_missing_fields(self):
+        """Defensive: ``approved`` / ``always`` / ``feedback`` may be
+        omitted by older nodes mid-rolling-upgrade; collector coerces
+        to safe defaults."""
+        c = _make_collector()
+        c._nodes["node-a"] = NodeSnapshot(
+            node_id="node-a",
+            server_url="http://a:8080",
+            workstreams={"ws1": {"id": "ws1", "name": "test", "state": "idle"}},
+        )
+        q: queue.Queue[dict] = queue.Queue()
+        c.register_listener(q)
+
+        c._apply_delta("node-a", {"type": "approval_resolved", "ws_id": "ws1"})
+
+        event = q.get_nowait()
+        assert event["approved"] is False
+        assert event["feedback"] == ""
+        assert event["always"] is False
+
     def test_apply_delta_health_changed(self):
         c = _make_collector()
         c._nodes["node-a"] = NodeSnapshot(

tests/test_coord_ui_approve_tools.py

@@ -464,3 +464,148 @@ def test_coord_budget_override_survives_wildcard_allow_policy() -> None:
     assert approved is True
     types = [e.get("type") for e in captured_events]
     assert "approve_request" in types, "Wildcard allow must not strip the budget-override prompt"
+
+
+# ---------------------------------------------------------------------------
+# Cluster-bus broadcast hooks — _broadcast_intent_verdict / _approval_resolved
+# ---------------------------------------------------------------------------
+
+
+class TestBroadcastIntentVerdict:
+    """``ConsoleCoordinatorUI._broadcast_intent_verdict`` overrides the
+    no-op base hook to push the verdict onto the cluster bus via
+    ``ClusterCollector.emit_console_ws_intent_verdict``. The far more
+    common path is the per-node ``WebUI`` override (covered in
+    test_webui_content.py); this lights up the rare coord-self path
+    (a coord that runs its own LLM judge).
+    """
+
+    def test_calls_collector_emit_with_ws_id_and_verdict(self) -> None:
+        ui = ConsoleCoordinatorUI(ws_id="coord-a", user_id="u1")
+        collector = MagicMock()
+        ConsoleCoordinatorUI._collector = collector
+        try:
+            verdict = {
+                "call_id": "c1",
+                "risk_level": "high",
+                "confidence": 0.91,
+            }
+            ui._broadcast_intent_verdict(verdict)
+            collector.emit_console_ws_intent_verdict.assert_called_once_with(
+                "coord-a",
+                verdict,
+            )
+        finally:
+            ConsoleCoordinatorUI._collector = None
+
+    def test_no_op_when_collector_unset(self) -> None:
+        ui = ConsoleCoordinatorUI(ws_id="coord-a", user_id="u1")
+        ConsoleCoordinatorUI._collector = None
+        # Doesn't raise.
+        ui._broadcast_intent_verdict({"call_id": "c1"})
+
+    def test_collector_exception_swallowed(self) -> None:
+        ui = ConsoleCoordinatorUI(ws_id="coord-a", user_id="u1")
+        collector = MagicMock()
+        collector.emit_console_ws_intent_verdict.side_effect = RuntimeError("boom")
+        ConsoleCoordinatorUI._collector = collector
+        try:
+            # Doesn't raise — collector failures are observational only.
+            ui._broadcast_intent_verdict({"call_id": "c1"})
+        finally:
+            ConsoleCoordinatorUI._collector = None
+
+
+class TestBroadcastApprovalResolved:
+    """``ConsoleCoordinatorUI._broadcast_approval_resolved`` overrides
+    the base hook to push the resolution onto the cluster bus via
+    ``ClusterCollector.emit_console_ws_approval_resolved``."""
+
+    def test_calls_collector_with_decision_fields(self) -> None:
+        ui = ConsoleCoordinatorUI(ws_id="coord-a", user_id="u1")
+        collector = MagicMock()
+        ConsoleCoordinatorUI._collector = collector
+        try:
+            ui._broadcast_approval_resolved(True, "lgtm", always=True)
+            collector.emit_console_ws_approval_resolved.assert_called_once_with(
+                "coord-a",
+                approved=True,
+                feedback="lgtm",
+                always=True,
+            )
+        finally:
+            ConsoleCoordinatorUI._collector = None
+
+    def test_normalises_none_feedback_to_empty_string(self) -> None:
+        ui = ConsoleCoordinatorUI(ws_id="coord-a", user_id="u1")
+        collector = MagicMock()
+        ConsoleCoordinatorUI._collector = collector
+        try:
+            ui._broadcast_approval_resolved(False, None)
+            collector.emit_console_ws_approval_resolved.assert_called_once_with(
+                "coord-a",
+                approved=False,
+                feedback="",
+                always=False,
+            )
+        finally:
+            ConsoleCoordinatorUI._collector = None
+
+    def test_no_op_when_collector_unset(self) -> None:
+        ui = ConsoleCoordinatorUI(ws_id="coord-a", user_id="u1")
+        ConsoleCoordinatorUI._collector = None
+        # Doesn't raise.
+        ui._broadcast_approval_resolved(True, None)
+
+    def test_collector_exception_swallowed(self) -> None:
+        ui = ConsoleCoordinatorUI(ws_id="coord-a", user_id="u1")
+        collector = MagicMock()
+        collector.emit_console_ws_approval_resolved.side_effect = RuntimeError("boom")
+        ConsoleCoordinatorUI._collector = collector
+        try:
+            # Doesn't raise.
+            ui._broadcast_approval_resolved(True, "ok")
+        finally:
+            ConsoleCoordinatorUI._collector = None
+
+
+class TestBroadcastApproveRequest:
+    """Coord-side override for the approve_request push. Same rationale
+    as the WebUI override — the coord-self path is rare today, but
+    parity keeps the override symmetric with the rest of the broadcast
+    family."""
+
+    def test_calls_collector_emit_with_ws_id_and_detail(self) -> None:
+        ui = ConsoleCoordinatorUI(ws_id="coord-a", user_id="u1")
+        collector = MagicMock()
+        ConsoleCoordinatorUI._collector = collector
+        try:
+            detail = {
+                "type": "approve_request",
+                "items": [{"call_id": "c1", "header": "tool x"}],
+                "judge_pending": True,
+            }
+            ui._broadcast_approve_request(detail)
+            collector.emit_console_ws_approve_request.assert_called_once_with(
+                "coord-a",
+                detail,
+            )
+        finally:
+            ConsoleCoordinatorUI._collector = None
+
+    def test_no_op_when_collector_unset(self) -> None:
+        ui = ConsoleCoordinatorUI(ws_id="coord-a", user_id="u1")
+        ConsoleCoordinatorUI._collector = None
+        # Doesn't raise.
+        ui._broadcast_approve_request({"items": []})
+
+    def test_collector_exception_swallowed(self) -> None:
+        ui = ConsoleCoordinatorUI(ws_id="coord-a", user_id="u1")
+        collector = MagicMock()
+        collector.emit_console_ws_approve_request.side_effect = RuntimeError("boom")
+        ConsoleCoordinatorUI._collector = collector
+        try:
+            # Doesn't raise.
+            ui._broadcast_approve_request({"items": []})
+        finally:
+            ConsoleCoordinatorUI._collector = None

tests/test_coordinator_adapter.py

@@ -378,13 +378,21 @@ class TestCoordinatorAdapterWorkerDispatch:
 
 
 class TestCoordinatorAdapterChildrenRegistry:
-    def test_emit_created_seeds_empty_children_set(self) -> None:
+    """Adapter-level integration with :class:`ChildrenRegistry`.
+
+    Pure-registry invariants (forward/reverse consistency, idempotent
+    merge, locking) live in ``test_children_registry.py``. These
+    tests cover the adapter's wiring: that ``emit_*`` paths drive the
+    registry correctly and that the snapshot-priming bridge between
+    a collector snapshot and the registry preserves merge semantics.
+    """
+
+    def test_emit_created_installs_parent(self) -> None:
         adapter, _ = _make_adapter()
         ws = _make_ws()
         adapter.emit_created(ws)
-        assert ws.id in adapter._children
-        assert adapter._children[ws.id] == set()
-        assert adapter._active_coords[ws.id] is ws.ui
+        assert adapter._registry.children_of(ws.id) == []
+        assert adapter._registry.ui_for(ws.id) is ws.ui
 
     def test_emit_rehydrated_calls_rebuild(self) -> None:
         adapter, _ = _make_adapter()
@@ -398,42 +406,36 @@ class TestCoordinatorAdapterChildrenRegistry:
         adapter.emit_rehydrated(ws)
         assert calls == [ws.id]
 
-    def test_emit_closed_clears_forward_and_reverse_indexes(self) -> None:
+    def test_emit_closed_uninstalls_parent_and_clears_children(self) -> None:
         adapter, _ = _make_adapter()
-        with adapter._children_lock:
-            adapter._merge_child_ids_locked("coord-a", ["child-a1", "child-a2"])
-            adapter._merge_child_ids_locked("coord-b", ["child-b1"])
-            adapter._active_coords["coord-a"] = object()
-            adapter._active_coords["coord-b"] = object()
+        adapter._registry.install("coord-a", object())
+        adapter._registry.install("coord-b", object())
+        adapter._registry.merge_children("coord-a", ["child-a1", "child-a2"])
+        adapter._registry.merge_children("coord-b", ["child-b1"])
 
         adapter.emit_closed("coord-a")
 
-        assert "coord-a" not in adapter._children
-        assert "coord-a" not in adapter._active_coords
-        assert "child-a1" not in adapter._child_to_coord
-        assert "child-a2" not in adapter._child_to_coord
+        assert adapter._registry.ui_for("coord-a") is None
+        assert adapter._registry.children_of("coord-a") == []
+        assert adapter._registry.parent_for("child-a1") is None
+        assert adapter._registry.parent_for("child-a2") is None
         # coord-b untouched
-        assert adapter._child_to_coord["child-b1"] == "coord-b"
-        assert "coord-b" in adapter._children
-
-    def test_merge_child_ids_locked_is_idempotent(self) -> None:
-        adapter, _ = _make_adapter()
-        with adapter._children_lock:
-            adapter._merge_child_ids_locked("coord-a", ["child-1"])
-            adapter._merge_child_ids_locked("coord-a", ["child-1"])
-        assert adapter._children["coord-a"] == {"child-1"}
-        assert adapter._child_to_coord == {"child-1": "coord-a"}
+        assert adapter._registry.parent_for("child-b1") == "coord-b"
+        assert adapter._registry.ui_for("coord-b") is not None
 
     def test_prime_children_from_snapshot_merges_without_overwriting(self) -> None:
+        # Snapshot priming now lives on ClusterChildSource (production
+        # path). The adapter no longer carries its own duplicate copy.
+        from turnstone.core.child_source import ClusterChildSource
+
         adapter, _ = _make_adapter()
-        # Seed one in-memory coord + one existing child
-        coord_ws = _make_ws()
-        coord_ws.id = "coord-a"
-        mgr = MagicMock()
-        mgr.list_all.return_value = [coord_ws]
-        adapter.attach(mgr)
-        with adapter._children_lock:
-            adapter._merge_child_ids_locked("coord-a", ["child-a1"])
+        adapter._registry.merge_children("coord-a", ["child-a1"])
+
+        source = ClusterChildSource(
+            collector=MagicMock(),
+            registry=adapter._registry,
+            parents_provider=lambda: ["coord-a"],
+        )
 
         snapshot = {
             "nodes": [
@@ -448,10 +450,13 @@ class TestCoordinatorAdapterChildrenRegistry:
                 },
             ],
         }
-        adapter._prime_children_from_snapshot(snapshot)
-        assert adapter._children["coord-a"] == {"child-a1", "child-a2"}
-        assert adapter._child_to_coord["child-a2"] == "coord-a"
-        assert "child-x" not in adapter._child_to_coord
+        source._prime_from_snapshot(snapshot)
+        assert set(adapter._registry.children_of("coord-a")) == {
+            "child-a1",
+            "child-a2",
+        }
+        assert adapter._registry.parent_for("child-a2") == "coord-a"
+        assert adapter._registry.parent_for("child-x") is None
 
 
 # ---------------------------------------------------------------------------
@@ -478,9 +483,7 @@ class TestCoordinatorAdapterDispatchChildEvent:
         coord_ws.id = coord_id
         recorder = _UIRecorder()
         coord_ws.ui = recorder  # type: ignore[assignment]
-        with adapter._children_lock:
-            adapter._children.setdefault(coord_id, set())
-            adapter._active_coords[coord_id] = recorder
+        adapter._registry.install(coord_id, recorder)
         adapter.attach(_StubManager(coord_ws))  # type: ignore[arg-type]
         return adapter, recorder, coord_ws
 
@@ -510,12 +513,11 @@ class TestCoordinatorAdapterDispatchChildEvent:
         assert payload["child_ws_id"] == "child-a1"
         assert payload["parent_ws_id"] == "coord-a"
         # Reverse index updated for subsequent cluster_state events.
-        assert adapter._child_to_coord["child-a1"] == "coord-a"
+        assert adapter._registry.parent_for("child-a1") == "coord-a"
 
     def test_dispatch_cluster_state_routes_via_reverse_index(self) -> None:
         adapter, recorder, _ = self._setup()
-        with adapter._children_lock:
-            adapter._merge_child_ids_locked("coord-a", ["child-a1"])
+        adapter._registry.merge_children("coord-a", ["child-a1"])
         adapter._dispatch_child_event(
             {
                 "type": "cluster_state",
@@ -533,8 +535,7 @@ class TestCoordinatorAdapterDispatchChildEvent:
 
     def test_dispatch_ws_closed_routes_to_parent_coord(self) -> None:
         adapter, recorder, _ = self._setup()
-        with adapter._children_lock:
-            adapter._merge_child_ids_locked("coord-a", ["child-a1"])
+        adapter._registry.merge_children("coord-a", ["child-a1"])
         adapter._dispatch_child_event(
             {"type": "ws_closed", "ws_id": "child-a1", "reason": "evicted"}
         )
@@ -548,8 +549,7 @@ class TestCoordinatorAdapterDispatchChildEvent:
         """perf-6: _enqueue_on_ui mutates the payload dict in place with
         the coord's ws_id so the browser can discriminate child events."""
         adapter, recorder, _ = self._setup()
-        with adapter._children_lock:
-            adapter._merge_child_ids_locked("coord-a", ["child-a1"])
+        adapter._registry.merge_children("coord-a", ["child-a1"])
         adapter._dispatch_child_event(
             {
                 "type": "cluster_state",
@@ -559,53 +559,160 @@ class TestCoordinatorAdapterDispatchChildEvent:
         )
         assert recorder.enqueued[0]["ws_id"] == "coord-a"
 
-    def test_dispatch_cluster_state_forwards_pending_approval_detail(self) -> None:
-        """The rich approval payload now rides on child_ws_state directly so
-        the browser can mutate liveBadgeCache without a separate live-bulk
-        fetch.  Drift here means the inline approve/deny buttons would
-        regress to chasing the dashboard cache (the load-storm pattern
-        Shape A is unwinding)."""
+    def test_dispatch_cluster_state_does_not_carry_pending_approval_detail(
+        self,
+    ) -> None:
+        """Stage 3 cleanup — the ``pending_approval_detail`` piggyback
+        on ``cluster_state`` is gone. Approval items now arrive via
+        bulk fetch (triggered by ``activity_state="approval"`` in the
+        browser); verdicts via ``child_ws_intent_verdict``; resolution
+        via ``child_ws_approval_resolved``. The state event carries
+        only state + activity_state — no detail field."""
         adapter, recorder, _ = self._setup()
-        with adapter._children_lock:
-            adapter._merge_child_ids_locked("coord-a", ["child-a1"])
-        detail = {
-            "items": [{"call_id": "c1", "header": "tool x"}],
-            "judge_pending": False,
-        }
+        adapter._registry.merge_children("coord-a", ["child-a1"])
         adapter._dispatch_child_event(
             {
                 "type": "cluster_state",
                 "ws_id": "child-a1",
                 "state": "running",
                 "activity_state": "approval",
-                "pending_approval_detail": detail,
             }
         )
         assert len(recorder.enqueued) == 1
         payload = recorder.enqueued[0]
         assert payload["type"] == "child_ws_state"
         assert payload["activity_state"] == "approval"
-        assert payload["pending_approval_detail"] == detail
+        assert "pending_approval_detail" not in payload
 
-    def test_dispatch_cluster_state_pending_approval_detail_none_passes_through(
-        self,
-    ) -> None:
-        """Missing pending_approval_detail (no approval pending, or pre-fix
-        node mid-rolling-upgrade) must forward as None — not raise, not
-        omit — so the browser's handleChildState treats it as "no SSE-
-        supplied detail, fall back to cached value"."""
+    def test_dispatch_intent_verdict_emits_child_ws_intent_verdict(self) -> None:
+        """Stage 3 Step 6 — explicit verdict events are re-emitted as
+        child_ws_intent_verdict on the parent's SSE so the tree UI
+        renders the risk pill without polling."""
         adapter, recorder, _ = self._setup()
-        with adapter._children_lock:
-            adapter._merge_child_ids_locked("coord-a", ["child-a1"])
+        adapter._registry.merge_children("coord-a", ["child-a1"])
+        verdict = {
+            "call_id": "c1",
+            "risk_level": "low",
+            "confidence": 0.92,
+            "recommendation": "approve",
+        }
         adapter._dispatch_child_event(
             {
-                "type": "cluster_state",
+                "type": "intent_verdict",
                 "ws_id": "child-a1",
-                "state": "running",
-                "activity_state": "tool",
+                "node_id": "node-1",
+                "verdict": verdict,
             }
         )
         assert len(recorder.enqueued) == 1
         payload = recorder.enqueued[0]
-        assert "pending_approval_detail" in payload
-        assert payload["pending_approval_detail"] is None
+        assert payload["type"] == "child_ws_intent_verdict"
+        assert payload["child_ws_id"] == "child-a1"
+        assert payload["parent_ws_id"] == "coord-a"
+        assert payload["node_id"] == "node-1"
+        assert payload["verdict"] == verdict
+
+    def test_dispatch_intent_verdict_unknown_child_drops(self) -> None:
+        adapter, recorder, _ = self._setup()
+        adapter._dispatch_child_event(
+            {
+                "type": "intent_verdict",
+                "ws_id": "ws-orphan",
+                "verdict": {"call_id": "c1"},
+            }
+        )
+        assert recorder.enqueued == []
+
+    def test_dispatch_approval_resolved_emits_child_ws_approval_resolved(
+        self,
+    ) -> None:
+        """Stage 3 Step 6 — paired with intent_verdict; clears the
+        pending-approval pill on the parent's tree UI in lockstep
+        with the actual decision."""
+        adapter, recorder, _ = self._setup()
+        adapter._registry.merge_children("coord-a", ["child-a1"])
+        adapter._dispatch_child_event(
+            {
+                "type": "approval_resolved",
+                "ws_id": "child-a1",
+                "node_id": "node-1",
+                "approved": True,
+                "feedback": "lgtm",
+                "always": False,
+            }
+        )
+        assert len(recorder.enqueued) == 1
+        payload = recorder.enqueued[0]
+        assert payload["type"] == "child_ws_approval_resolved"
+        assert payload["child_ws_id"] == "child-a1"
+        assert payload["parent_ws_id"] == "coord-a"
+        assert payload["approved"] is True
+        assert payload["feedback"] == "lgtm"
+        assert payload["always"] is False
+
+    def test_dispatch_approval_resolved_coerces_missing_fields(self) -> None:
+        """Older nodes mid-rolling-upgrade may omit approved / always /
+        feedback; dispatch coerces to safe defaults."""
+        adapter, recorder, _ = self._setup()
+        adapter._registry.merge_children("coord-a", ["child-a1"])
+        adapter._dispatch_child_event({"type": "approval_resolved", "ws_id": "child-a1"})
+        assert len(recorder.enqueued) == 1
+        payload = recorder.enqueued[0]
+        assert payload["approved"] is False
+        assert payload["feedback"] == ""
+        assert payload["always"] is False
+
+    def test_dispatch_approval_resolved_unknown_child_drops(self) -> None:
+        """Symmetric to the intent_verdict drop test — events for
+        ws_ids the registry doesn't know about silently drop instead
+        of fanning out to a parent that has no business seeing them."""
+        adapter, recorder, _ = self._setup()
+        adapter._dispatch_child_event(
+            {
+                "type": "approval_resolved",
+                "ws_id": "ws-orphan",
+                "approved": True,
+            },
+        )
+        assert recorder.enqueued == []
+
+    def test_dispatch_approve_request_emits_child_ws_approve_request(
+        self,
+    ) -> None:
+        """Push path for the initial approval items — eliminates the
+        bulk-fetch race that left the coord row stuck on a loading
+        placeholder when the bulk fetch landed in the gap between
+        _emit_state(ATTENTION) and approve_tools setting _pending_approval."""
+        adapter, recorder, _ = self._setup()
+        adapter._registry.merge_children("coord-a", ["child-a1"])
+        detail = {
+            "type": "approve_request",
+            "items": [{"call_id": "c1", "header": "tool x"}],
+            "judge_pending": True,
+        }
+        adapter._dispatch_child_event(
+            {
+                "type": "approve_request",
+                "ws_id": "child-a1",
+                "node_id": "node-1",
+                "detail": detail,
+            },
+        )
+        assert len(recorder.enqueued) == 1
+        payload = recorder.enqueued[0]
+        assert payload["type"] == "child_ws_approve_request"
+        assert payload["child_ws_id"] == "child-a1"
+        assert payload["parent_ws_id"] == "coord-a"
+        assert payload["node_id"] == "node-1"
+        assert payload["detail"] == detail
+
+    def test_dispatch_approve_request_unknown_child_drops(self) -> None:
+        adapter, recorder, _ = self._setup()
+        adapter._dispatch_child_event(
+            {
+                "type": "approve_request",
+                "ws_id": "ws-orphan",
+                "detail": {"items": []},
+            },
+        )
+        assert recorder.enqueued == []

tests/test_coordinator_page.py

@@ -79,9 +79,10 @@ def test_coordinator_js_exposes_inline_approval_helpers():
     assert "function submitChildApproval" in body or "submitChildApproval(" in body
     # The shared approve POST helper (parameterized for child ws_ids)
     assert "function approveWorkstream" in body or "approveWorkstream(" in body
-    # The urgent live-bulk fetch option that fires on activity_state
-    # transitions in/out of "approval"
-    assert "{ urgent: true }" in body or "urgent: true" in body
+    # The 409 stale-call_id retry path uses invalidateLiveBadge +
+    # scheduleLiveFetch (Stage 3 cleanup removed the urgent flag —
+    # cache invalidation makes the TTL gate fall through naturally).
+    assert "invalidateLiveBadge(targetWsId)" in body
     # Server-side payload field — drift here means the JS reads stale keys
     assert "pending_approval_detail" in body
     # Reconnect parity (chunk 4): the SSE re-open handler must drop
@@ -90,10 +91,10 @@ def test_coordinator_js_exposes_inline_approval_helpers():
     # can't render zombie approve/deny buttons on a row whose
     # approval was resolved during the gap. The implementation
     # iterates the cache and deletes only !permanent entries —
-    # asserting the literal Map iteration form keeps a refactor
-    # back to liveBadgeCache.clear() (which would re-pay 403s on
-    # every reconnect for denied ids) from sneaking in.
-    assert "liveBadgeCache.delete" in body
+    # asserting the literal helper call keeps a refactor back to
+    # _liveBadgeCacheClear() (which would re-pay 403s on every
+    # reconnect for denied ids) from sneaking in.
+    assert "_liveBadgeCacheDelete" in body
     # Edge-case matrix sentinel labels — POLICY-BLOCKED renders when
     # an item has error set + needs_approval=False (server-side
     # tool policy already blocked the call); "(judge unavailable)"
@@ -113,15 +114,19 @@ def test_coordinator_js_exposes_inline_approval_helpers():
     # coord-self ws_id (the coord lives on the console process).
     # Children live on cluster nodes and 404 without the prefix.
     assert "/v1/api/route/workstreams/" in body
-    # Late-judge polling — the LLM judge runs async on the child
-    # node and never pushes a signal that reaches the coord, so
-    # the row's pending_approval_detail with judge_pending=true
-    # would freeze on heuristic verdicts forever without this
-    # poll loop. The poller is GLOBAL (not per-row) so off-screen
-    # rows still refresh — a per-row poller's scheduleLiveFetch
-    # call short-circuits on non-visible rows, leaving them stuck.
-    assert "_maybeStartJudgePoll" in body
-    assert "_judgePollTick" in body
+    # Late-arriving LLM judge verdicts — Stage 3 Step 5 promoted
+    # ``intent_verdict`` and ``approval_resolved`` to first-class
+    # cluster-bus event types, so the coord adapter dispatches them
+    # as ``child_ws_intent_verdict`` / ``child_ws_approval_resolved``
+    # on the parent's SSE stream. The browser handlers write
+    # directly to liveBadgeCache (bypassing scheduleLiveFetch's
+    # visibility gate cleanly) so off-screen rows pick up verdicts
+    # without polling. Replaced the old ``_judgePollTick`` 90-second
+    # global poll loop and its visibility-gate-bypass workaround.
+    assert "handleChildIntentVerdict" in body
+    assert "handleChildApprovalResolved" in body
+    assert "child_ws_intent_verdict" in body
+    assert "child_ws_approval_resolved" in body
     # Reload parity for the coord-self approval gate: init() must
     # consume the authoritative GET /workstreams snapshot's
     # pending_approval_detail so a freshly opened tab can render
@@ -154,18 +159,19 @@ def test_coordinator_js_exposes_inline_approval_helpers():
     assert "callOutcomes" in body
 
 
-def test_coordinator_js_handle_child_state_reads_sse_pending_approval_detail():
-    """Lock the Shape A behavior change: child_ws_state SSE events now
-    carry ``pending_approval_detail`` directly so the browser mutates
-    ``liveBadgeCache`` without firing an urgent live-bulk fetch on
-    every activity_state transition into/out of approval.  A refactor
-    that re-introduces the urgent-fetch path on routine transitions
-    (or drops the SSE-source merge guard in flushLiveFetches) would
-    re-open the load-storm pattern this PR is fixing.
+def test_coordinator_js_handle_child_state_no_longer_reads_sse_pending_approval_detail():
+    """Stage 3 cleanup — ``pending_approval_detail`` is no longer
+    piggybacked on child_ws_state events. Approval items now arrive
+    via bulk fetch on the activity_state="approval" transition;
+    verdicts via the explicit ``child_ws_intent_verdict`` event class;
+    resolution via ``child_ws_approval_resolved``. A refactor that
+    re-introduces the piggyback would silently re-open the
+    duplicate-path race the dedicated event classes were added to
+    eliminate.
 
     Structural assertions (regex against multi-line source) — symbol-
     presence alone wouldn't catch a guard that keeps the names but
-    inverts the comparison or drops the ``prev.live`` check.  This
+    inverts the comparison or drops the ``prev.live`` check. This
     codebase has no JS test framework, so locking the guard's shape
     here is the next-best thing to a behavioral test."""
     import re
@@ -176,26 +182,46 @@ def test_coordinator_js_handle_child_state_reads_sse_pending_approval_detail():
     )
     body = coord_js.read_text(encoding="utf-8")
 
-    # handleChildState now reads the SSE-supplied detail.
-    assert "ev.pending_approval_detail" in body
-    # The pre-fix urgent-fetch on activity_state transitions is
-    # gone (the 409 retry path keeps its own ``{ urgent: true }``
-    # for stale-call_id refresh — that's a different scenario).
+    # The piggyback read is gone from handleChildState. (The string
+    # may still appear elsewhere — e.g. handleChildIntentVerdict
+    # reading from cache, or comments — but never as ``ev.pending_approval_detail``.)
+    assert "ev.pending_approval_detail" not in body
+    # The pre-fix urgent-fetch on activity_state transitions is gone.
     assert "enteredApproval" not in body
     assert "leftApproval" not in body
+    # ``pendingApproval`` flag derivation must check BOTH state and
+    # activity_state. The worker thread can fire the state transition
+    # to "attention" before approve_tools updates activity_state, so
+    # checking only activity_state misses children that legitimately
+    # need approval. Pin the disjunction so the regression doesn't
+    # silently re-introduce.
+    assert re.search(
+        r'existing\.state\s*===\s*"attention"\s*\|\|\s*'
+        r'existing\.activity_state\s*===\s*"approval"',
+        body,
+    ), (
+        "handleChildState must derive pendingApproval from "
+        "(state==='attention' || activity_state==='approval')"
+    )
 
     # SSE-authoritative window constant is defined and used.
     assert re.search(r"\bconst\s+SSE_AUTHORITATIVE_MS\s*=\s*\d+", body), (
         "SSE_AUTHORITATIVE_MS constant must be defined as a numeric literal"
     )
 
-    # handleChildState writes sseUpdatedAt = Date.now() into the cache
-    # entry it sets.  This is the SSE-source tag; without it, the
-    # merge guard in flushLiveFetches has nothing to gate on.
+    # SSE writers tag entries with sseUpdatedAt: Date.now() so the
+    # merge guard in flushLiveFetches preserves them against stale
+    # bulk-fetch responses. handleChildState only stamps when it
+    # AUTHORITATIVELY clears the detail (off-approval transition);
+    # writers that stamp unconditionally are intent_verdict (verdict
+    # stamp), approval_resolved (clear), and the optimistic-clear
+    # path in submitChildApproval. Pinning the literal Date.now()
+    # call keeps a refactor that drops the SSE-source tag entirely
+    # from sneaking in.
     assert re.search(
         r"sseUpdatedAt:\s*Date\.now\(\)",
         body,
-    ), "handleChildState must write sseUpdatedAt: Date.now() onto liveBadgeCache entries"
+    ), "Critical SSE writers must stamp sseUpdatedAt: Date.now()"
 
     # flushLiveFetches' merge guard structure: SSE-set pending_approval
     # / _detail wins over a stale bulk-poll snapshot when (live) AND

tests/test_session_manager.py

@@ -19,7 +19,10 @@ import threading
 import time
 from dataclasses import dataclass
 from datetime import UTC, datetime
-from typing import Any
+from typing import TYPE_CHECKING, Any
+
+if TYPE_CHECKING:
+    from collections.abc import Callable
 from unittest.mock import MagicMock
 
 import pytest
@@ -1368,3 +1371,127 @@ class TestSessionManagerWithStateWriter:
         assert "running" not in ws_writes, (
             f"set_state after close enqueued through buffer: {ws_writes}"
         )
+
+
+# ---------------------------------------------------------------------------
+# Multi-subscriber observer — subscribe_to_state / unsubscribe_from_state
+# ---------------------------------------------------------------------------
+
+
+class TestStateSubscribers:
+    """Multi-subscriber observer for ``set_state``.
+
+    Used by the CLI's background-attention notifier and by
+    ``SameNodeChildSource``. Subscribe / unsubscribe must be safe under
+    concurrent dispatch, and dispatch must not skip / repeat callbacks
+    when subscribers register or unregister mid-iteration.
+    """
+
+    def test_subscribe_fires_on_set_state(self) -> None:
+        mgr, _, _ = _make_manager()
+        ws = mgr.create(user_id="u1", name="ws", skill=None)
+        events: list[tuple[str, str]] = []
+
+        def cb(ws_id: str, state: WorkstreamState) -> None:
+            events.append((ws_id, state.value))
+
+        mgr.subscribe_to_state(cb)
+        mgr.set_state(ws.id, WorkstreamState.RUNNING)
+        assert events == [(ws.id, "running")]
+
+    def test_unsubscribe_stops_firing(self) -> None:
+        mgr, _, _ = _make_manager()
+        ws = mgr.create(user_id="u1", name="ws", skill=None)
+        events: list[str] = []
+
+        def cb(_ws_id: str, state: WorkstreamState) -> None:
+            events.append(state.value)
+
+        mgr.subscribe_to_state(cb)
+        mgr.unsubscribe_from_state(cb)
+        mgr.set_state(ws.id, WorkstreamState.RUNNING)
+        assert events == []
+
+    def test_unsubscribe_unknown_is_noop(self) -> None:
+        mgr, _, _ = _make_manager()
+        # Doesn't raise.
+        mgr.unsubscribe_from_state(lambda *_: None)
+
+    def test_multiple_subscribers_fire_in_registration_order(self) -> None:
+        mgr, _, _ = _make_manager()
+        ws = mgr.create(user_id="u1", name="ws", skill=None)
+        order: list[int] = []
+
+        def make(i: int) -> Callable[[str, WorkstreamState], None]:
+            def cb(_ws_id: str, _state: WorkstreamState) -> None:
+                order.append(i)
+
+            return cb
+
+        mgr.subscribe_to_state(make(1))
+        mgr.subscribe_to_state(make(2))
+        mgr.subscribe_to_state(make(3))
+        mgr.set_state(ws.id, WorkstreamState.IDLE)
+        assert order == [1, 2, 3]
+
+    def test_subscriber_exception_does_not_block_others(self) -> None:
+        mgr, _, _ = _make_manager()
+        ws = mgr.create(user_id="u1", name="ws", skill=None)
+        survived: list[str] = []
+
+        def boom(*_: Any) -> None:
+            raise RuntimeError("subscriber crash")
+
+        def good(_ws_id: str, state: WorkstreamState) -> None:
+            survived.append(state.value)
+
+        mgr.subscribe_to_state(boom)
+        mgr.subscribe_to_state(good)
+        mgr.set_state(ws.id, WorkstreamState.RUNNING)
+        assert survived == ["running"]
+
+    @pytest.mark.parametrize("n_threads", [10, 50])
+    def test_concurrent_subscribe(self, n_threads: int) -> None:
+        """Subscribe from many threads; all callbacks land in the list.
+
+        Validates the lock around mutation — without it the underlying
+        list.append could lose entries under contention.
+        """
+        mgr, _, _ = _make_manager()
+        callbacks = [lambda *_, i=i: None for i in range(n_threads)]
+        threads = [threading.Thread(target=mgr.subscribe_to_state, args=(cb,)) for cb in callbacks]
+        for t in threads:
+            t.start()
+        for t in threads:
+            t.join()
+        # Snapshot under the lock to read the count safely.
+        with mgr._state_subscribers_lock:
+            assert len(mgr._state_subscribers) == n_threads
+
+    def test_subscribe_during_dispatch_does_not_corrupt_iteration(self) -> None:
+        """A subscriber that calls subscribe_to_state during its own
+        callback must not affect the in-flight dispatch (snapshot
+        isolation). This is the bug-1 invariant: mutation during
+        iteration can't shift the iterator's index because dispatch
+        iterates a snapshot, not the live list.
+        """
+        mgr, _, _ = _make_manager()
+        ws = mgr.create(user_id="u1", name="ws", skill=None)
+        fired: list[str] = []
+
+        def late(_ws_id: str, state: WorkstreamState) -> None:
+            fired.append("late:" + state.value)
+
+        def first(_ws_id: str, state: WorkstreamState) -> None:
+            fired.append("first:" + state.value)
+            mgr.subscribe_to_state(late)  # mid-dispatch addition
+
+        mgr.subscribe_to_state(first)
+        mgr.set_state(ws.id, WorkstreamState.RUNNING)
+        # ``late`` was added during dispatch but the snapshot was
+        # already frozen — so it doesn't fire on this round.
+        assert fired == ["first:running"]
+        # Next round it does fire, in registration order after first.
+        fired.clear()
+        mgr.set_state(ws.id, WorkstreamState.IDLE)
+        assert fired == ["first:idle", "late:idle"]

tests/test_webui_content.py

@@ -157,20 +157,17 @@ class TestContentAccumulation:
         assert len(idle_events[0]["content"]) <= _MAX_TURN_CONTENT_CHARS + 1024
 
 
-class TestPendingApprovalDetailGate:
-    """The Shape A SSE plumbing carries ``pending_approval_detail`` on the
-    ``ws_state`` event so the coord tree UI can render inline approve/deny
-    buttons in lockstep with the activity_state transition.  The gate
-    (``if self._pending_approval is not None``) keeps the per-broadcast
-    serializer cost off the common no-approval-pending path — these tests
-    lock both branches down."""
+class TestPendingApprovalDetailNotPiggybacked:
+    """Stage 3 cleanup — ``pending_approval_detail`` is no longer
+    piggybacked on ``ws_state`` events. Approval items now arrive via
+    bulk fetch when the coord tree's reducer sees the
+    ``activity_state="approval"`` transition; verdicts via the explicit
+    ``intent_verdict`` event class; resolution via
+    ``approval_resolved``. These tests lock the no-piggyback contract
+    down so a future regression doesn't silently re-introduce the
+    duplicated path."""
 
     def test_state_broadcast_omits_field_when_no_approval_pending(self):
-        """Common case: no approval pending → field absent from event so the
-        per-broadcast verdict-cache deepcopy in
-        ``serialize_pending_approval_detail`` never runs.  A regression
-        that drops the gate would silently 10x the cost of every state
-        broadcast in the steady state."""
         ui = _make_ui()
         assert ui._pending_approval is None
         ui._broadcast_state("running")
@@ -180,17 +177,12 @@ class TestPendingApprovalDetailGate:
         assert len(running_events) == 1
         assert "pending_approval_detail" not in running_events[0]
 
-    def test_state_broadcast_includes_field_when_approval_pending(self):
-        """When an approval is pending the broadcast must carry the rich
-        payload — the coord tree UI reads it directly to render inline
-        approve/deny buttons.  Without this, a coord browser would have
-        to chase a separate ``cluster/ws/live`` fetch on every
-        activity_state transition (the load-storm pattern Shape A is
-        unwinding)."""
+    def test_state_broadcast_omits_field_even_when_approval_pending(self):
+        """The piggyback is gone: even when ``_pending_approval`` is set,
+        the state broadcast must NOT carry ``pending_approval_detail``.
+        The browser triggers a bulk fetch off the
+        ``activity_state="approval"`` transition to get the items."""
         ui = _make_ui()
-        # Mirror the shape ``pause_for_approval`` writes (session_ui_base
-        # lines 576-580) — items with call_id + header is the minimum
-        # the serializer needs to project.
         ui._pending_approval = {
             "type": "approve_request",
             "items": [
@@ -209,21 +201,9 @@ class TestPendingApprovalDetailGate:
         events = _drain_global()
         attn = [e for e in events if e.get("state") == "attention"]
         assert len(attn) == 1
-        # Field present and structurally sound — the serializer's
-        # full shape is covered by tests/test_session_ui_base.py;
-        # here we only need to confirm the gate fires and the
-        # serializer's output is what lands on the event.
-        assert "pending_approval_detail" in attn[0]
-        detail = attn[0]["pending_approval_detail"]
-        assert detail is not None
-        assert detail.get("items")
-        assert detail["items"][0]["call_id"] == "c1"
+        assert "pending_approval_detail" not in attn[0]
 
-    def test_field_cleared_after_approval_resolves(self):
-        """Once ``_pending_approval`` is cleared, subsequent state
-        broadcasts must drop the field again — without this, the
-        browser would render stale approve/deny buttons until the
-        next bulk-poll TTL window expired."""
+    def test_field_stays_absent_after_approval_resolves(self):
         ui = _make_ui()
         ui._pending_approval = {
             "type": "approve_request",
@@ -231,7 +211,7 @@ class TestPendingApprovalDetailGate:
             "judge_pending": False,
         }
         ui._broadcast_state("attention")
-        _drain_global()  # discard the with-detail event
+        _drain_global()
 
         ui._pending_approval = None
         ui._broadcast_state("running")
@@ -239,3 +219,115 @@ class TestPendingApprovalDetailGate:
         running = [e for e in events if e.get("state") == "running"]
         assert len(running) == 1
         assert "pending_approval_detail" not in running[0]
+
+
+class TestBroadcastIntentVerdict:
+    """Producer-side coverage for ``WebUI._broadcast_intent_verdict``.
+
+    The collector-side test (``test_apply_delta_intent_verdict_*`` in
+    test_console.py) covers consumption; this pins the event shape the
+    producer puts on the global queue. A field rename or missed key
+    here would slip past the consumer test because the consumer reads
+    via ``data.get(...)``.
+    """
+
+    def test_pushes_intent_verdict_event_to_global_queue(self):
+        ui = _make_ui()
+        verdict = {
+            "call_id": "c1",
+            "risk_level": "low",
+            "confidence": 0.92,
+            "recommendation": "approve",
+            "reasoning": "tool reads only",
+        }
+        ui._broadcast_intent_verdict(verdict)
+
+        events = _drain_global()
+        assert len(events) == 1
+        ev = events[0]
+        assert ev["type"] == "intent_verdict"
+        assert ev["ws_id"] == "ws-test"
+        assert ev["verdict"] == verdict
+
+    def test_no_op_when_global_queue_unset(self):
+        WebUI._global_queue = None
+        ui = _make_ui()
+        # Doesn't raise.
+        ui._broadcast_intent_verdict({"call_id": "c1"})
+
+    def test_queue_full_swallowed(self):
+        # Force a tiny queue then fill it so the next put_nowait
+        # raises queue.Full — the broadcast must absorb it without
+        # propagating (matches _broadcast_state's queue.Full handling).
+        WebUI._global_queue = queue.Queue(maxsize=1)
+        WebUI._global_queue.put_nowait({"sentinel": True})
+        ui = _make_ui()
+        # Doesn't raise.
+        ui._broadcast_intent_verdict({"call_id": "c1"})
+
+
+class TestBroadcastApprovalResolved:
+    """Producer-side coverage for ``WebUI._broadcast_approval_resolved``."""
+
+    def test_pushes_approval_resolved_event_to_global_queue(self):
+        ui = _make_ui()
+        ui._broadcast_approval_resolved(True, "lgtm", always=False)
+
+        events = _drain_global()
+        assert len(events) == 1
+        ev = events[0]
+        assert ev["type"] == "approval_resolved"
+        assert ev["ws_id"] == "ws-test"
+        assert ev["approved"] is True
+        assert ev["feedback"] == "lgtm"
+        assert ev["always"] is False
+
+    def test_normalises_none_feedback_to_empty_string(self):
+        ui = _make_ui()
+        ui._broadcast_approval_resolved(False, None)
+
+        events = _drain_global()
+        assert events[0]["feedback"] == ""
+        assert events[0]["approved"] is False
+        assert events[0]["always"] is False
+
+    def test_always_kwarg_propagates(self):
+        ui = _make_ui()
+        ui._broadcast_approval_resolved(True, "ok", always=True)
+        events = _drain_global()
+        assert events[0]["always"] is True
+
+    def test_no_op_when_global_queue_unset(self):
+        WebUI._global_queue = None
+        ui = _make_ui()
+        # Doesn't raise.
+        ui._broadcast_approval_resolved(True, None)
+
+
+class TestBroadcastApproveRequest:
+    """Producer-side coverage for ``WebUI._broadcast_approve_request`` —
+    push path for the initial approval items so a coord parent's tree
+    UI can render the inline approve/deny block immediately without
+    waiting for a bulk-fetch round-trip."""
+
+    def test_pushes_approve_request_event_to_global_queue(self):
+        ui = _make_ui()
+        detail = {
+            "type": "approve_request",
+            "items": [{"call_id": "c1", "header": "tool x"}],
+            "judge_pending": True,
+        }
+        ui._broadcast_approve_request(detail)
+
+        events = _drain_global()
+        assert len(events) == 1
+        ev = events[0]
+        assert ev["type"] == "approve_request"
+        assert ev["ws_id"] == "ws-test"
+        assert ev["detail"] == detail
+
+    def test_no_op_when_global_queue_unset(self):
+        WebUI._global_queue = None
+        ui = _make_ui()
+        # Doesn't raise.
+        ui._broadcast_approve_request({"items": []})

turnstone/cli.py

@@ -1242,7 +1242,7 @@ def main() -> None:
                 )
                 sys.stderr.flush()
 
-    manager._on_state_change = _bg_attention_notify
+    manager.subscribe_to_state(_bg_attention_notify)
 
     # Print banner
     print(f"\n{bold('Chat')} with {cyan(model)}")

turnstone/console/collector.py

@@ -12,16 +12,22 @@ import asyncio
 import contextlib
 import json
 import logging
-import queue
 import random
 import threading
 import time
 from dataclasses import dataclass, field
 from typing import TYPE_CHECKING, Any
 
+if TYPE_CHECKING:
+    import queue
+
 import httpx
 import httpx_sse
 
+from turnstone.core.session_ui_base import (
+    _CRITICAL_EVENT_TYPES,
+    _put_with_priority,
+)
 from turnstone.core.workstream import WorkstreamKind
 
 if TYPE_CHECKING:
@@ -157,11 +163,19 @@ class ClusterCollector:
         log.info("ClusterCollector stopped")
 
     def _fanout(self, event: dict[str, Any]) -> None:
-        """Copy an event to all registered SSE listener queues."""
+        """Copy an event to all registered SSE listener queues.
+
+        Critical event types (verdicts, approval resolutions, child
+        lifecycle) evict one oldest queue entry to make room rather
+        than dropping themselves on a full queue. Best-effort events
+        (state ticks, activity) drop as before. See
+        :data:`turnstone.core.session_ui_base._CRITICAL_EVENT_TYPES`
+        for the canonical list.
+        """
+        critical = event.get("type") in _CRITICAL_EVENT_TYPES
         with self._listeners_lock:
             for q in self._listeners:
-                with contextlib.suppress(queue.Full):
-                    q.put_nowait(event)
+                _put_with_priority(q, event, critical=critical)
 
     # -- auth helpers --------------------------------------------------------
 
@@ -498,7 +512,6 @@ class ClusterCollector:
                         "kind": WorkstreamKind.from_raw(new_w.get("kind")),
                         "parent_ws_id": new_w.get("parent_ws_id"),
                         "activity_state": new_w.get("activity_state", ""),
-                        "pending_approval_detail": new_w.get("pending_approval_detail"),
                     }
                 )
             old_name = old_ws.get("title", "") or old_ws.get("name", "")
@@ -558,18 +571,6 @@ class ClusterCollector:
                         ws["kind"] = data["kind"]
                     if "parent_ws_id" in data:
                         ws["parent_ws_id"] = data["parent_ws_id"]
-                    # ``pending_approval_detail`` overwrites (no
-                    # ``ws.get`` fallback): the node's broadcast gate
-                    # on ``_pending_approval is not None`` means the
-                    # field is absent from ``data`` exactly when no
-                    # approval is pending — falling back to the cached
-                    # value would resurrect a stale detail after the
-                    # approval resolved.  Without this assignment the
-                    # cached ``node.workstreams`` dict served by
-                    # ``get_node_detail`` / ``get_snapshot`` between
-                    # reconciliations would render stale approve/deny
-                    # buttons on closed approvals.
-                    ws["pending_approval_detail"] = data.get("pending_approval_detail")
                     pending_events.append(
                         {
                             "type": "cluster_state",
@@ -581,7 +582,6 @@ class ClusterCollector:
                             "kind": WorkstreamKind.from_raw(ws.get("kind")),
                             "parent_ws_id": ws.get("parent_ws_id"),
                             "activity_state": ws.get("activity_state", ""),
-                            "pending_approval_detail": data.get("pending_approval_detail"),
                         }
                     )
 
@@ -648,6 +648,60 @@ class ClusterCollector:
                     ws["name"] = name
                 pending_events.append({"type": "ws_rename", "ws_id": ws_id, "name": name})
 
+            elif etype == "intent_verdict":
+                # Pure pass-through for LLM judge verdicts. The node's
+                # WebUI._broadcast_intent_verdict pushes onto the
+                # global queue (Stage 3 Step 5); the cluster fan-out
+                # forwards to subscribers (e.g. CoordinatorAdapter,
+                # which dispatches to the parent's coord SSE as
+                # ``child_ws_intent_verdict``).
+                ws_id = data.get("ws_id", "")
+                if ws_id:
+                    pending_events.append(
+                        {
+                            "type": "intent_verdict",
+                            "ws_id": ws_id,
+                            "node_id": node_id,
+                            "verdict": data.get("verdict") or {},
+                        }
+                    )
+
+            elif etype == "approval_resolved":
+                # Pure pass-through for approve/deny resolutions. Pairs
+                # with ``intent_verdict`` above so the parent
+                # coordinator's tree UI can clear the pending-approval
+                # pill the moment the user decides, rather than waiting
+                # for the subsequent state-change piggyback.
+                ws_id = data.get("ws_id", "")
+                if ws_id:
+                    pending_events.append(
+                        {
+                            "type": "approval_resolved",
+                            "ws_id": ws_id,
+                            "node_id": node_id,
+                            "approved": bool(data.get("approved", False)),
+                            "feedback": data.get("feedback", "") or "",
+                            "always": bool(data.get("always", False)),
+                        }
+                    )
+
+            elif etype == "approve_request":
+                # Push path for the initial approval items. Eliminates
+                # the bulk-fetch race that otherwise leaves the coord
+                # tree stuck on a loading placeholder when the bulk
+                # fetch lands in the gap between the state transition
+                # to ATTENTION and ``_pending_approval`` being set.
+                ws_id = data.get("ws_id", "")
+                if ws_id:
+                    pending_events.append(
+                        {
+                            "type": "approve_request",
+                            "ws_id": ws_id,
+                            "node_id": node_id,
+                            "detail": data.get("detail") or {},
+                        }
+                    )
+
             elif etype == "health_changed":
                 # Update the health dict's backend status in-place
                 bstatus = data.get("backend_status", "")
@@ -1217,3 +1271,70 @@ class ClusterCollector:
                 return
             entry["name"] = name
         self._fanout({"type": "ws_rename", "ws_id": ws_id, "name": name})
+
+    def emit_console_ws_intent_verdict(self, ws_id: str, verdict: dict[str, Any]) -> None:
+        """Fan an LLM intent-judge verdict for a console-pseudo-node ws.
+
+        Gives coord-spawned approval flows a first-class cluster-bus
+        event so the parent's tree UI can render the risk pill +
+        verdict result without polling.
+        ``CoordinatorAdapter._dispatch_child_event`` re-emits these as
+        ``child_ws_intent_verdict`` for the parent coordinator's SSE
+        stream. The dispatch path filters by registry membership;
+        downstream subscribers tolerate verdicts for ws_ids they don't
+        own (silently drop), so we skip the membership pre-check that
+        would otherwise add a lock acquisition per emit on a path
+        that fires once per tool-call during heuristic+LLM judging.
+        """
+        self._fanout(
+            {
+                "type": "intent_verdict",
+                "ws_id": ws_id,
+                "node_id": self.CONSOLE_PSEUDO_NODE_ID,
+                "verdict": verdict,
+            }
+        )
+
+    def emit_console_ws_approval_resolved(
+        self,
+        ws_id: str,
+        *,
+        approved: bool,
+        feedback: str = "",
+        always: bool = False,
+    ) -> None:
+        """Fan an ``approval_resolved`` decision for a console-pseudo-node ws.
+
+        Paired with :meth:`emit_console_ws_intent_verdict` so the
+        coord tree UI clears the pending-approval pill in lockstep
+        with the actual decision. Same lock-skip rationale as the
+        intent-verdict emit above.
+        """
+        self._fanout(
+            {
+                "type": "approval_resolved",
+                "ws_id": ws_id,
+                "node_id": self.CONSOLE_PSEUDO_NODE_ID,
+                "approved": approved,
+                "feedback": feedback,
+                "always": always,
+            }
+        )
+
+    def emit_console_ws_approve_request(self, ws_id: str, detail: dict[str, Any]) -> None:
+        """Fan an ``approve_request`` payload for a console-pseudo-node ws.
+
+        Push path for the initial approval items so a coord parent's
+        tree UI can render the inline approve/deny block immediately
+        without a bulk-fetch round-trip. ``CoordinatorAdapter._dispatch_child_event``
+        re-emits as ``child_ws_approve_request`` for the parent
+        coordinator's SSE stream.
+        """
+        self._fanout(
+            {
+                "type": "approve_request",
+                "ws_id": ws_id,
+                "node_id": self.CONSOLE_PSEUDO_NODE_ID,
+                "detail": detail,
+            }
+        )

turnstone/console/coordinator_adapter.py

@@ -15,17 +15,17 @@ for the storage-seeded children rebuild.
 
 from __future__ import annotations
 
-import queue
-import threading
 from typing import TYPE_CHECKING, Any
 
 from turnstone.core import session_worker
 from turnstone.core.adapters._ui_cleanup import cleanup_session_ui
+from turnstone.core.child_source import ClusterChildSource
+from turnstone.core.children_registry import ChildrenRegistry
 from turnstone.core.log import get_logger
 from turnstone.core.workstream import Workstream, WorkstreamKind, WorkstreamState
 
 if TYPE_CHECKING:
-    from collections.abc import Callable, Iterable
+    from collections.abc import Callable
 
     from turnstone.console.collector import ClusterCollector
     from turnstone.console.coordinator_ui import ConsoleCoordinatorUI
@@ -56,38 +56,25 @@ class CoordinatorAdapter:
         # the manager to the adapter's ``__init__`` — break the cycle
         # with a setter called from console startup.
         self._manager: SessionManager | None = None
-        # Per-coordinator known-child ws_id set. Populated lazily on
-        # create/open from storage and updated live as the cluster fan-out
-        # thread sees ws_created events with matching parent_ws_id.
-        # Closed / deleted children stay in the registry so the tree UI
-        # can keep rendering them grayed out; the authoritative render
-        # path reads storage for state. Bounded by eventual coordinator
-        # close/eviction.
-        self._children: dict[str, set[str]] = {}
-        # Reverse index for O(1) child → coord lookup on every cluster
-        # event. Without this, every cluster event incurs a linear scan
-        # over every coordinator's child set while holding the fan-out
-        # lock — a hot-path tax that scales with both active coordinators
-        # and their retained-history depth.
-        self._child_to_coord: dict[str, str] = {}
-        self._children_lock = threading.Lock()
-        # Cluster-event fan-out: subscribes to the ClusterCollector's
-        # listener channel, filters by known child ws_ids, and re-emits
-        # child_ws_* events on the matching coordinator's UI. Configured
-        # lazily via ``start_child_event_fanout(collector)`` from the
-        # console lifespan once both the manager and collector exist.
-        self._collector_queue: queue.Queue[dict[str, Any]] | None = None
-        self._fanout_thread: threading.Thread | None = None
-        self._fanout_stop = threading.Event()
-        # Coord-ws-id → UI map for the fan-out dispatch path. Read
-        # and written under ``self._children_lock`` alongside the
-        # forward/reverse child maps. (Previously the value was
-        # ``(user_id, ui)`` with a copy-on-write dict swap so the
-        # dispatch could read it lock-free — but _dispatch_child_event
-        # already re-validates the parent under _children_lock anyway,
-        # so the lock-free snapshot was premature. The user_id half is
-        # also dead after a46dab1 dropped row-level ownership gates.)
-        self._active_coords: dict[str, Any] = {}
+        # Children registry — universal parent → children + reverse
+        # lookup primitive. Lifted from inline data on this adapter to
+        # ``turnstone.core.children_registry.ChildrenRegistry`` (Stage 3
+        # Step 1) so the same primitive can serve interactive
+        # workstreams when they gain spawn capability. The dispatch
+        # path (`_dispatch_child_event`) calls into the registry
+        # atomically; everything else delegates through the legacy
+        # method names which still exist as thin shims for the
+        # cluster-routing + cleanup callers.
+        self._registry = ChildrenRegistry()
+        # Cross-node child events arrive via ``ClusterChildSource``
+        # (Stage 3 Step 2): a strategy that subscribes to the
+        # collector's listener channel and runs a daemon thread that
+        # drains the queue, pushing each event to the sink. The sink
+        # is :meth:`_dispatch_child_event` so the existing translation
+        # logic (cluster_state → child_ws_state, etc.) stays in one
+        # place. Constructed lazily by ``start_child_event_fanout`` so
+        # the collector reference is available.
+        self._child_source: ClusterChildSource | None = None
 
     def attach(self, manager: SessionManager) -> None:
         """Late-bind the owning :class:`SessionManager`.
@@ -109,7 +96,7 @@ class CoordinatorAdapter:
         # needs the empty forward/presence entries so
         # ``_dispatch_child_event`` recognises this coordinator when
         # its first child is spawned.
-        self._install_coord_registry(ws)
+        self._registry.install(ws.id, ws.ui)
         self._fanout_console_ws_created(ws)
 
     def emit_rehydrated(self, ws: Workstream) -> None:
@@ -117,21 +104,10 @@ class CoordinatorAdapter:
         # it from storage after the registry seed so a ``ws_created``
         # for an already-spawned child that fires mid-rebuild merges
         # cleanly.
-        self._install_coord_registry(ws)
+        self._registry.install(ws.id, ws.ui)
         self._rebuild_children_registry(ws.id)
         self._fanout_console_ws_created(ws)
 
-    def _install_coord_registry(self, ws: Workstream) -> None:
-        """Seed the children registry + presence map for ``ws``.
-
-        Shared by ``emit_created`` and ``emit_rehydrated`` — the
-        difference between the two is purely whether we then rebuild
-        from storage.
-        """
-        with self._children_lock:
-            self._children.setdefault(ws.id, set())
-            self._active_coords[ws.id] = ws.ui
-
     def _fanout_console_ws_created(self, ws: Workstream) -> None:
         try:
             self._collector.emit_console_ws_created(
@@ -197,14 +173,11 @@ class CoordinatorAdapter:
         # toast only fire for real-node (interactive) ws_closed events.
         del reason, name
         # Drop the coordinator's children-registry entries AND its
-        # presence slot. Mirrors the eviction/close paths from the old
-        # CoordinatorManager (which did the same under _children_lock
-        # + _lock respectively). A plain _children.pop without clearing
-        # the reverse index would leak every evicted coordinator's
-        # child→parent pointers forever.
-        with self._children_lock:
-            self._pop_coord_registry_locked(ws_id)
-            self._active_coords.pop(ws_id, None)
+        # presence slot. A plain pop without clearing the reverse
+        # index would leak every evicted coordinator's child→parent
+        # pointers forever — :meth:`ChildrenRegistry.uninstall`
+        # handles forward set + reverse entries + presence atomically.
+        self._registry.uninstall(ws_id)
         try:
             self._collector.emit_console_ws_closed(ws_id)
         except Exception:
@@ -356,25 +329,6 @@ class CoordinatorAdapter:
     # Children registry
     # ------------------------------------------------------------------
 
-    def _merge_child_ids_locked(self, coord_ws_id: str, child_ids: Iterable[str]) -> None:
-        """Merge ``child_ids`` into ``coord_ws_id``'s forward + reverse maps.
-
-        Caller MUST hold ``self._children_lock``. Idempotent — re-adding
-        an existing child is a no-op (the reverse-index pointer is
-        already correct). Empty / falsy entries in ``child_ids`` are
-        skipped.
-
-        Sole write-path for bulk registry updates so
-        ``_rebuild_children_registry`` (storage-seeded) and
-        ``_prime_children_from_snapshot`` (collector-seeded) agree on
-        ordering and reverse-index invariants.
-        """
-        existing = self._children.setdefault(coord_ws_id, set())
-        for cid in child_ids:
-            if cid and cid not in existing:
-                existing.add(cid)
-                self._child_to_coord[cid] = coord_ws_id
-
     def _rebuild_children_registry(self, coord_ws_id: str) -> None:
         """Populate ``self._children[coord_ws_id]`` from storage.
 
@@ -445,174 +399,88 @@ class CoordinatorAdapter:
             if not child_id:
                 continue
             child_ids.append(child_id)
-        with self._children_lock:
-            self._merge_child_ids_locked(coord_ws_id, child_ids)
-
-    def _coord_for_child(self, child_ws_id: str) -> str | None:
-        """Reverse-lookup: which coordinator owns this child ws_id?
-
-        O(1) via the ``_child_to_coord`` reverse index. Cluster events
-        fire on every token tick across the cluster; a linear scan here
-        turned into a hot-path tax as the retained-history set grew.
-        """
-        with self._children_lock:
-            return self._child_to_coord.get(child_ws_id)
+        self._registry.merge_children(coord_ws_id, child_ids)
 
     def children_snapshot(self, coord_ws_id: str) -> list[str]:
         """Return a snapshot of the coordinator's direct child ws_ids.
 
-        Used by ``stop_cascade`` to iterate children without holding the
-        registry lock during the per-child HTTP dispatch. A mutation
-        racing with the snapshot (child spawned mid-cascade) either
-        lands before the snapshot and gets cancelled, or lands after
-        and is out of scope for this batch — both outcomes are safe.
-        Returns an empty list for unknown coordinators.
+        Used by ``stop_cascade`` to iterate children without holding
+        the registry lock during the per-child HTTP dispatch. A
+        mutation racing with the snapshot (child spawned mid-cascade)
+        either lands before (cancelled) or after (out of scope for
+        this batch) — both safe. Returns an empty list for unknown
+        coordinators.
         """
-        with self._children_lock:
-            child_set = self._children.get(coord_ws_id)
-            return list(child_set) if child_set else []
-
-    def _pop_coord_registry_locked(self, coord_ws_id: str) -> None:
-        """Remove a coordinator's forward set + reverse-index entries.
-
-        Caller MUST hold ``self._children_lock``. Used by close /
-        eviction paths so stale coordinators don't leak registry
-        entries. No-op if the coordinator is unknown.
-        """
-        child_set = self._children.pop(coord_ws_id, None)
-        if child_set is None:
-            return
-        for cid in child_set:
-            # Defensive: only clear the reverse entry if it still points
-            # at THIS coordinator. If a child has since been reassigned
-            # (unusual but possible on schema changes), we don't want to
-            # orphan the new owner's entry.
-            if self._child_to_coord.get(cid) == coord_ws_id:
-                self._child_to_coord.pop(cid, None)
+        return self._registry.children_of(coord_ws_id)
 
     # ------------------------------------------------------------------
     # Cluster-event fan-out thread
     # ------------------------------------------------------------------
 
     def start_child_event_fanout(self, collector: ClusterCollector) -> None:
-        """Subscribe to cluster events and start the filter + re-emit thread.
+        """Subscribe to cluster events via :class:`ClusterChildSource`.
 
-        Idempotent — calling twice is a no-op (already-started fan-out
-        thread stays). Called once from the console lifespan after both
-        the collector and the session manager are constructed.
+        Idempotent — already-started ChildSource stays. Called once
+        from the console lifespan after both the collector and the
+        session manager are constructed.
         """
-        if self._fanout_thread is not None and self._fanout_thread.is_alive():
+        if self._child_source is not None:
             return
         self._collector = collector
-        self._collector_queue = queue.Queue(maxsize=1000)
-        # Ensure the "console" pseudo-node exists in the snapshot map so
-        # emit_console_ws_* calls from create / close / open land on a
-        # real node entry the snapshot will surface.
+        # Coord-specific transport setup: the "console" pseudo-node
+        # must exist in the snapshot map BEFORE any
+        # ``emit_console_ws_*`` calls (and before the snapshot is
+        # taken inside ``ChildSource.start``) so those emits land on
+        # a real node entry the snapshot surfaces.
         collector.ensure_console_pseudo_node()
-        # Register with the collector — use the existing listener channel
-        # the browser SSE fan-out uses; the collector treats our queue as
-        # just another subscriber.
-        snapshot = collector.get_snapshot_and_register(self._collector_queue)
-        # Prime the child registry from the snapshot so a coordinator
-        # that opens right after a console restart sees already-live
-        # children without waiting for the next ``ws_state`` tick to
-        # discover them via the fan-out path.
-        self._prime_children_from_snapshot(snapshot)
-        # Seed the pseudo-node with any coordinators already loaded in
-        # memory when the collector binds. Prevents a race where early
-        # creates happened before the collector was wired up and their
-        # rows never showed on the snapshot.
-        mgr = self._manager
-        if mgr is not None:
-            for ws in mgr.list_all():
-                try:
-                    collector.emit_console_ws_created(
-                        ws.id,
-                        name=ws.name,
-                        user_id=ws.user_id or "",
-                        kind=WorkstreamKind.COORDINATOR.value,
-                        state=ws.state.value,
-                        parent_ws_id=None,
-                    )
-                except Exception:
-                    log.debug(
-                        "coord_adapter.collector_seed_failed ws=%s",
-                        ws.id[:8],
-                        exc_info=True,
-                    )
-        self._fanout_stop.clear()
-        t = threading.Thread(
-            target=self._fanout_loop,
-            name="coord-adapter-child-fanout",
-            daemon=True,
-        )
-        self._fanout_thread = t
-        t.start()
-
-    def _prime_children_from_snapshot(self, snapshot: dict[str, Any]) -> None:
-        """Populate ``_children`` + ``_child_to_coord`` from a collector snapshot.
-
-        The snapshot's per-node workstreams carry ``parent_ws_id``. For
-        every workstream whose parent is an in-memory coordinator,
-        record the child so the fan-out filter sees it immediately.
-        """
-        nodes = snapshot.get("nodes", []) if isinstance(snapshot, dict) else []
-        if not nodes:
-            return
         mgr = self._manager
         if mgr is None:
             raise RuntimeError(
                 "CoordinatorAdapter: manager not attached — call attach(mgr) after construction"
             )
-        by_parent: dict[str, list[str]] = {}
-        known = {ws.id for ws in mgr.list_all()}
-        for node in nodes:
-            for entry in node.get("workstreams", []) or []:
-                parent = entry.get("parent_ws_id") or ""
-                child_id = entry.get("id") or ""
-                if not parent or not child_id or parent not in known:
-                    continue
-                by_parent.setdefault(parent, []).append(child_id)
-        if not by_parent:
-            return
-        with self._children_lock:
-            for parent, kids in by_parent.items():
-                self._merge_child_ids_locked(parent, kids)
+        # Build + start the strategy. The sink is
+        # :meth:`_dispatch_child_event` so cluster events flow through
+        # the same translation path that synthesises ``child_ws_*``
+        # payloads for the parent's UI.
+        source = ClusterChildSource(
+            collector=collector,
+            registry=self._registry,
+            parents_provider=lambda: [ws.id for ws in mgr.list_all()],
+        )
+        source.start(sink=self._dispatch_child_event)
+        self._child_source = source
+        # Seed the pseudo-node with any coordinators already loaded in
+        # memory when the collector binds. Prevents a race where early
+        # creates happened before the collector was wired up and their
+        # rows never showed on the snapshot. (Coord-specific — interactive
+        # has no analogous pseudo-node.)
+        for ws in mgr.list_all():
+            try:
+                collector.emit_console_ws_created(
+                    ws.id,
+                    name=ws.name,
+                    user_id=ws.user_id or "",
+                    kind=WorkstreamKind.COORDINATOR.value,
+                    state=ws.state.value,
+                    parent_ws_id=None,
+                )
+            except Exception:
+                log.debug(
+                    "coord_adapter.collector_seed_failed ws=%s",
+                    ws.id[:8],
+                    exc_info=True,
+                )
 
     def shutdown(self) -> None:
-        """Stop the fan-out thread and unregister from the collector.
+        """Stop the ChildSource and unregister from the collector.
 
         Safe to call multiple times; idempotent. Invoked from the
         console lifespan teardown so SSE listener queues don't leak.
         """
-        self._fanout_stop.set()
-        t = self._fanout_thread
-        q = self._collector_queue
-        coll = self._collector
-        self._fanout_thread = None
-        self._collector_queue = None
-        if coll is not None and q is not None:
-            try:
-                coll.unregister_listener(q)
-            except Exception:
-                log.debug("coord_adapter.unregister_listener_failed", exc_info=True)
-        if t is not None:
-            t.join(timeout=2.0)
-
-    def _fanout_loop(self) -> None:
-        """Drain collector events, filter by known children, dispatch."""
-        q = self._collector_queue
-        if q is None:
-            return
-        while not self._fanout_stop.is_set():
-            try:
-                event = q.get(timeout=1.0)
-            except queue.Empty:
-                continue
-            try:
-                self._dispatch_child_event(event)
-            except Exception:
-                log.debug("coord_adapter.fanout.dispatch_failed", exc_info=True)
+        src = self._child_source
+        self._child_source = None
+        if src is not None:
+            src.shutdown()
 
     def _dispatch_child_event(self, event: dict[str, Any]) -> None:
         """Match a cluster event to a coordinator and re-emit on its UI.
@@ -622,10 +490,12 @@ class CoordinatorAdapter:
         - ``ws_created`` with ``parent_ws_id`` matching an in-memory
           coordinator → add to registry + re-emit as
           ``child_ws_created``.
-        - ``cluster_state`` / ``ws_closed`` / ``ws_rename`` whose
-          ``ws_id`` is in any coordinator's known-children registry →
-          re-emit as ``child_ws_state`` / ``child_ws_closed`` /
-          ``child_ws_rename``.
+        - ``cluster_state`` / ``ws_closed`` / ``ws_rename`` /
+          ``intent_verdict`` / ``approval_resolved`` whose ``ws_id``
+          is in any coordinator's known-children registry → re-emit as
+          ``child_ws_state`` / ``child_ws_closed`` /
+          ``child_ws_rename`` / ``child_ws_intent_verdict`` /
+          ``child_ws_approval_resolved``.
 
         Events for ws_ids we don't own silently drop — the filter lives
         on the server so each coordinator's SSE stream stays small.
@@ -639,23 +509,17 @@ class CoordinatorAdapter:
             parent = event.get("parent_ws_id") or ""
             if not parent:
                 return
-            # Presence check + registry mutation under the same lock:
-            # a concurrent close()/eviction can pop the entry between
-            # the check and the mutation, after which a bare setdefault
-            # would resurrect the entry — leaking the registry key and
-            # enqueuing onto the closed coordinator's UI. Trusted-team
-            # posture (#400 / a46dab1) means no per-event tenant gate
-            # here; scope-level auth at the SSE endpoint is the only
-            # boundary.
-            with self._children_lock:
-                coord_ui = self._active_coords.get(parent)
-                if coord_ui is None:
-                    return
-                existing = self._children.setdefault(parent, set())
-                if ws_id in existing:
-                    return
-                existing.add(ws_id)
-                self._child_to_coord[ws_id] = parent
+            # Atomic check-and-route under the registry's lock: a
+            # concurrent close()/eviction can pop the parent's entry
+            # between the presence check and the mutation, so the two
+            # must happen together. ``add_child`` returns the parent's
+            # UI on success or None on (a) parent not installed, or
+            # (b) duplicate child. Trusted-team posture (#400 / a46dab1)
+            # means no per-event tenant gate here; scope-level auth at
+            # the SSE endpoint is the only boundary.
+            coord_ui = self._registry.add_child(parent, ws_id)
+            if coord_ui is None:
+                return
             payload = {
                 "type": "child_ws_created",
                 "ws_id": ws_id,
@@ -668,8 +532,15 @@ class CoordinatorAdapter:
             _enqueue_on_ui(coord_ui, parent, payload)
             return
 
-        if etype in ("cluster_state", "ws_closed", "ws_rename"):
-            coord_id = self._coord_for_child(ws_id)
+        if etype in (
+            "cluster_state",
+            "ws_closed",
+            "ws_rename",
+            "intent_verdict",
+            "approval_resolved",
+            "approve_request",
+        ):
+            coord_id = self._registry.parent_for(ws_id)
             if coord_id is None:
                 return
             mgr = self._manager
@@ -684,14 +555,13 @@ class CoordinatorAdapter:
                     "state": event.get("state", ""),
                     "tokens": event.get("tokens", 0),
                     "node_id": event.get("node_id", ""),
-                    # activity_state lets the JS detect approval-state
-                    # transitions; pending_approval_detail rides on
-                    # the same event so the browser can mutate
-                    # liveBadgeCache directly and render inline
-                    # approve/deny buttons in lockstep with the
-                    # transition, no separate dashboard refetch.
+                    # ``activity_state`` lets the JS detect approval
+                    # transitions and trigger a bulk fetch for the
+                    # initial detail. The detail itself no longer
+                    # piggybacks here (Stage 3 cleanup) — verdicts
+                    # arrive via the explicit ``intent_verdict`` event
+                    # class and resolution via ``approval_resolved``.
                     "activity_state": event.get("activity_state", ""),
-                    "pending_approval_detail": event.get("pending_approval_detail"),
                 }
             elif etype == "ws_closed":
                 child_event = {
@@ -700,13 +570,52 @@ class CoordinatorAdapter:
                     "parent_ws_id": coord_id,
                     "reason": event.get("reason", ""),
                 }
-            else:  # ws_rename
+            elif etype == "ws_rename":
                 child_event = {
                     "type": "child_ws_rename",
                     "child_ws_id": ws_id,
                     "parent_ws_id": coord_id,
                     "name": event.get("name", ""),
                 }
+            elif etype == "intent_verdict":
+                # Per-coord re-emit of an explicit verdict event so
+                # the tree UI can render the risk pill + verdict
+                # result without polling. The corresponding
+                # ``cluster_state`` event no longer carries the
+                # detail (the piggyback was removed end-to-end);
+                # the bulk fetch on initial approval entry plus this
+                # explicit event class are the only carriers.
+                child_event = {
+                    "type": "child_ws_intent_verdict",
+                    "child_ws_id": ws_id,
+                    "parent_ws_id": coord_id,
+                    "node_id": event.get("node_id", ""),
+                    "verdict": event.get("verdict") or {},
+                }
+            elif etype == "approval_resolved":
+                child_event = {
+                    "type": "child_ws_approval_resolved",
+                    "child_ws_id": ws_id,
+                    "parent_ws_id": coord_id,
+                    "node_id": event.get("node_id", ""),
+                    "approved": bool(event.get("approved", False)),
+                    "feedback": event.get("feedback", "") or "",
+                    "always": bool(event.get("always", False)),
+                }
+            else:  # approve_request
+                # Push path for the initial approval items —
+                # eliminates the bulk-fetch race that previously left
+                # the coord row stuck on a loading placeholder when
+                # the bulk fetch landed in the gap between the state
+                # transition to ATTENTION and ``_pending_approval``
+                # being set inside ``approve_tools``.
+                child_event = {
+                    "type": "child_ws_approve_request",
+                    "child_ws_id": ws_id,
+                    "parent_ws_id": coord_id,
+                    "node_id": event.get("node_id", ""),
+                    "detail": event.get("detail") or {},
+                }
             _enqueue_on_ui(owning_ws.ui, coord_id, child_event)
 
 

turnstone/console/coordinator_ui.py

@@ -181,6 +181,81 @@ class ConsoleCoordinatorUI(SessionUIBase):
         with self._ws_lock:
             self._last_broadcast_activity = current
 
+    def _broadcast_intent_verdict(self, verdict: dict[str, Any]) -> None:
+        """Fan an LLM intent-judge verdict to the cluster collector.
+
+        Stage 3 Step 5 — overrides the no-op base hook so a coord
+        workstream that produces its own verdict (rare — coord agents
+        don't typically run the LLM judge) gets a first-class
+        cluster-bus event for the parent's tree UI. The far more
+        common path is a CHILD workstream firing its verdict on a
+        real node; that path goes through ``WebUI._broadcast_intent_verdict``
+        → global SSE → collector ``_apply_delta``.
+        """
+        collector = ConsoleCoordinatorUI._collector
+        if collector is None:
+            return
+        try:
+            collector.emit_console_ws_intent_verdict(self.ws_id, verdict)
+        except Exception:
+            log.debug(
+                "coord_ui.intent_verdict_fanout_failed ws=%s",
+                self.ws_id,
+                exc_info=True,
+            )
+
+    def _broadcast_approval_resolved(
+        self,
+        approved: bool,
+        feedback: str | None = None,
+        *,
+        always: bool = False,
+    ) -> None:
+        """Fan an ``approval_resolved`` decision to the cluster collector.
+
+        Stage 3 Step 5 — paired with :meth:`_broadcast_intent_verdict`.
+        Same rationale: coord-direct approvals are rare; the typical
+        path is a child workstream resolving on its node, with that
+        node's WebUI broadcasting through the global queue.
+        """
+        collector = ConsoleCoordinatorUI._collector
+        if collector is None:
+            return
+        try:
+            collector.emit_console_ws_approval_resolved(
+                self.ws_id,
+                approved=approved,
+                feedback=feedback or "",
+                always=always,
+            )
+        except Exception:
+            log.debug(
+                "coord_ui.approval_resolved_fanout_failed ws=%s",
+                self.ws_id,
+                exc_info=True,
+            )
+
+    def _broadcast_approve_request(self, detail: dict[str, Any]) -> None:
+        """Fan an ``approve_request`` payload to the cluster collector.
+
+        Push path for the initial approval items. Same rationale as
+        the other two broadcast hooks: coord-self approvals are rare
+        (the LLM judge isn't wired on the console coord today), but
+        the override exists for symmetry and lights up the same path
+        a future coord-self gate would use.
+        """
+        collector = ConsoleCoordinatorUI._collector
+        if collector is None:
+            return
+        try:
+            collector.emit_console_ws_approve_request(self.ws_id, detail)
+        except Exception:
+            log.debug(
+                "coord_ui.approve_request_fanout_failed ws=%s",
+                self.ws_id,
+                exc_info=True,
+            )
+
     def on_state_change(self, state: str) -> None:
         # Flow state transitions through the unified SessionManager so
         # the storage write + adapter emit_state fan-out stay in lockstep

turnstone/console/server.py

@@ -846,8 +846,20 @@ async def _fetch_live_block(
             live = {k: entry.get(k) for k in _CLUSTER_WS_LIVE_KEYS if k in entry}
             # Derived field — kept in lockstep with
             # _coordinator_live_snapshot so both origins produce the
-            # same keys.
-            live["pending_approval"] = live.get("activity_state") == "approval"
+            # same keys. ``state="attention"`` is the canonical signal;
+            # ``activity_state="approval"`` is set inside approve_tools
+            # AFTER the state transition fires, so a bulk fetch that
+            # races with that window can see state=attention and
+            # activity_state="" simultaneously. A non-null
+            # ``pending_approval_detail`` is also a definitive signal
+            # (the serializer only emits non-None when ``_pending_approval``
+            # is set on the UI). Any of the three flips this true; the
+            # frontend reducer mirrors the same disjunction.
+            live["pending_approval"] = (
+                live.get("activity_state") == "approval"
+                or entry.get("state") == "attention"
+                or live.get("pending_approval_detail") is not None
+            )
             return live
     return None
 

turnstone/console/static/coordinator/coordinator.js

@@ -1730,7 +1730,7 @@
         // state) so a user lacking admin.cluster.inspect doesn't
         // pay one 403 per denied id on every reconnect.
         for (const [id, c] of liveBadgeCache) {
-          if (!c || !c.permanent) liveBadgeCache.delete(id);
+          if (!c || !c.permanent) _liveBadgeCacheDelete(id);
         }
       }
     };
@@ -2038,6 +2038,21 @@
       case "child_ws_rename":
         handleChildRename(ev);
         break;
+      // Stage 3 Step 7 — explicit verdict + approval-resolved events
+      // arrive without piggybacking on cluster_state, so verdicts that
+      // land while a child is in `attention` (no state transition)
+      // reach the parent's tree UI promptly. Reducer writes
+      // liveBadgeCache directly (bypassing scheduleLiveFetch) so
+      // off-screen rows update too.
+      case "child_ws_intent_verdict":
+        handleChildIntentVerdict(ev);
+        break;
+      case "child_ws_approval_resolved":
+        handleChildApprovalResolved(ev);
+        break;
+      case "child_ws_approve_request":
+        handleChildApproveRequest(ev);
+        break;
       // wait_for_workstream observability (#14) — the worker thread
       // can block up to 600s inside the tool; these events drive a
       // sidebar indicator so operators see the coordinator is alive.
@@ -2050,12 +2065,40 @@
       case "wait_ended":
         handleWaitEnded(ev);
         break;
+      // Server-side SSE queue depth — debug aid surfaced in the
+      // status bar so operators can spot a backed-up tab (slow
+      // consumer / browser throttling) without devtools.
+      case "_queue_stats":
+        handleQueueStats(ev);
+        break;
       default:
         // Unknown event type — ignore silently.
         break;
     }
   }
 
+  // ------------------------------------------------------------------
+  // SSE queue depth indicator
+  // ------------------------------------------------------------------
+  //
+  // The server emits a periodic ``_queue_stats`` event carrying the
+  // per-tab listener queue depth. Surface in the coord status bar
+  // with color escalation (default → warn at >50% → danger at >80%)
+  // so a backed-up consumer is glanceable. The element starts
+  // hidden; first sample reveals it.
+  function handleQueueStats(ev) {
+    const el = document.getElementById("coord-sb-sse-queue");
+    if (!el) return;
+    const depth = typeof ev.depth === "number" ? ev.depth : 0;
+    const max = typeof ev.max === "number" && ev.max > 0 ? ev.max : 500;
+    el.hidden = false;
+    el.textContent = "queue " + depth + "/" + max;
+    el.classList.remove("warn", "danger");
+    const ratio = depth / max;
+    if (ratio > 0.8) el.classList.add("danger");
+    else if (ratio > 0.5) el.classList.add("warn");
+  }
+
   // ------------------------------------------------------------------
   // wait_for_workstream progress indicator (#14)
   // ------------------------------------------------------------------
@@ -2154,6 +2197,37 @@
   const childrenState = new Map();
   // ws_id -> {live: <dict>, fetched: <ms>} for the 5s TTL live-badge cache.
   const liveBadgeCache = new Map();
+  // Incrementally-maintained set of ws_ids currently flagged
+  // ``pending_approval`` in the cache. Updated at every cache mutation
+  // via ``_liveBadgeCacheSet`` / ``_liveBadgeCacheDelete`` /
+  // ``_liveBadgeCacheClear`` below so the sidebar pending-count read
+  // is O(1) instead of an O(N) walk over the cache on every render
+  // (caught by /review perf-1).
+  const pendingApprovalIds = new Set();
+  // The helpers reach into the Map's prototype directly so a future
+  // edit can't accidentally rewrite the body's ``liveBadgeCache.set``
+  // into the helper name and create infinite recursion (already
+  // happened once when bulk-replacing call sites — caught at runtime
+  // as "InternalError: too much recursion").
+  const _mapSet = Map.prototype.set;
+  const _mapDelete = Map.prototype.delete;
+  const _mapClear = Map.prototype.clear;
+  function _liveBadgeCacheSet(id, entry) {
+    _mapSet.call(liveBadgeCache, id, entry);
+    if (entry && entry.live && entry.live.pending_approval) {
+      pendingApprovalIds.add(id);
+    } else {
+      pendingApprovalIds.delete(id);
+    }
+  }
+  function _liveBadgeCacheDelete(id) {
+    _mapDelete.call(liveBadgeCache, id);
+    pendingApprovalIds.delete(id);
+  }
+  function _liveBadgeCacheClear() {
+    _mapClear.call(liveBadgeCache);
+    pendingApprovalIds.clear();
+  }
   // ws_ids currently visible in the viewport — only these trigger
   // live-fetch on SSE state changes.  Populated by an
   // IntersectionObserver attached to each rendered .ch-row so a
@@ -2280,23 +2354,22 @@
       }
     }
     row.appendChild(meta);
-    // Inline approve/deny block \u2014 shown only when the live block
-    // carries pending_approval_detail (the rich payload added by the
-    // server-side dashboard projection).  A "\u2691 approval" badge alone
-    // means the child is in attention state but the rich detail
-    // hasn't arrived on the cache yet \u2014 a rare cross-version race
-    // (e.g. a node mid-rolling-upgrade emitted ws_state without
-    // pending_approval_detail before this PR landed).  The next SSE
-    // tick or the 5s TTL bulk-poll catches up and re-renders.
-    if (cached && cached.live && cached.live.pending_approval_detail) {
+    // Inline approve/deny block \u2014 the detail arrives via the bulk
+    // fetch triggered by handleChildState off the activity_state="approval"
+    // transition. Verdicts that land later arrive via
+    // child_ws_intent_verdict and update the cached detail in place;
+    // resolution arrives via child_ws_approval_resolved and clears it.
+    // While the bulk fetch is in flight (~250ms debounce + ~100ms HTTP)
+    // we render a loading placeholder so the row keeps its height
+    // stable AND so a screen reader has a labelled region to land on
+    // \u2014 without it the operator sees a "demand for action" badge
+    // with no actionable content.
+    if (cached && cached.live && cached.live.pending_approval) {
       const detail = cached.live.pending_approval_detail;
-      const block = renderApprovalBlock(child, detail);
+      const block = detail
+        ? renderApprovalBlock(child, detail)
+        : renderApprovalPlaceholder(child);
       if (block) row.appendChild(block);
-      // Late-arriving LLM judge: see comment above
-      // _maybeStartJudgePoll for the why. Hooks into a single
-      // global poller (not per-row) so off-screen rows still
-      // refresh and one bulk request covers every pending row.
-      _maybeStartJudgePoll();
     }
     // Recent auto-approves — tools that bypassed the operator gate
     // (skill ``allowed_tools`` allowlist / blanket / admin policy /
@@ -2370,78 +2443,6 @@
     return pill;
   }
 
-  // Late-judge polling — single global timer driving a bulk
-  // re-fetch of every row whose `pending_approval_detail.judge_pending`
-  // is true and not every item has a `judge_verdict` yet.  Necessary
-  // because the LLM judge runs async on the child node and never
-  // pushes a signal that reaches the coord; without polling the
-  // operator would stare at a "heuristic"-tier pill forever.
-  //
-  // Why a global poller instead of per-row:
-  //  1. Per-row would invalidate the cache and call scheduleLiveFetch
-  //     for off-screen rows, but scheduleLiveFetch returns early on
-  //     non-visible ids — leaving the cache empty AND no fetch in
-  //     flight.  Result: off-screen rows stuck on stale heuristic.
-  //  2. A single bulk request covers every pending row in one
-  //     round-trip; the per-row design would issue N microtask-
-  //     batched fetches that share the bulk path anyway.
-  const JUDGE_POLL_INTERVAL_MS = 2000;
-  // Cap by total wall-clock time, not attempts. Real LLM judges
-  // (esp. with reasoning effort) can exceed 30s — a 6-attempt /
-  // 12s cap was prematurely giving up.
-  const JUDGE_POLL_MAX_DURATION_MS = 90_000;
-  let judgePollTimer = null;
-  let judgePollStartedAt = 0;
-
-  function _maybeStartJudgePoll() {
-    if (judgePollTimer !== null) return; // already polling
-    judgePollStartedAt = Date.now();
-    judgePollTimer = setTimeout(_judgePollTick, JUDGE_POLL_INTERVAL_MS);
-  }
-
-  function _judgePollTick() {
-    judgePollTimer = null;
-    if (Date.now() - judgePollStartedAt > JUDGE_POLL_MAX_DURATION_MS) {
-      // Failed / timed-out judge — give up so we don't poll forever.
-      // Operator can hit the Refresh button to force a fresh fetch.
-      return;
-    }
-    // Walk the full childrenState — not just visible — so off-screen
-    // rows still get refreshed.  scheduleLiveFetch's visibility gate
-    // exists to keep idle rows from burning round-trips; here we
-    // explicitly want every pending-judge row in the next bulk
-    // regardless of viewport.
-    let stillPending = false;
-    for (const [wsId, entry] of childrenState) {
-      if (TERMINAL_CHILD_STATES.has(entry.state)) continue;
-      const cached = liveBadgeCache.get(wsId);
-      if (!cached || !cached.live) continue;
-      const detail = cached.live.pending_approval_detail;
-      if (!detail || !detail.judge_pending) continue;
-      const items = Array.isArray(detail.items) ? detail.items : [];
-      const allHaveJudge =
-        items.length > 0 && items.every((it) => it.judge_verdict);
-      if (allHaveJudge) continue;
-      // Bypass scheduleLiveFetch entirely (skips visibility + TTL
-      // gates) by adding to pendingLiveIds directly. flushLiveFetches
-      // will batch every pending row into one request.
-      if (WS_ID_RE.test(wsId)) {
-        pendingLiveIds.add(wsId);
-        stillPending = true;
-      }
-    }
-    if (!stillPending) return; // every verdict landed — done
-    // Cancel any debounce that's still pending so our flush runs
-    // now instead of waiting for it. Then flush directly so the
-    // bulk request fires before the next tick re-arms.
-    if (liveBadgeFlushTimer !== null) {
-      clearTimeout(liveBadgeFlushTimer);
-      liveBadgeFlushTimer = null;
-    }
-    flushLiveFetches();
-    judgePollTimer = setTimeout(_judgePollTick, JUDGE_POLL_INTERVAL_MS);
-  }
-
   // Build the inline approval block: severity pill, intent summary +
   // judge reasoning, and approve/deny buttons.  Returns a DOM node or
   // null if the detail is unusable (defensive \u2014 server is supposed to
@@ -2534,6 +2535,37 @@
     return sub;
   }
 
+  // Loading-state placeholder rendered while the bulk fetch is
+  // in-flight. Same outer ``.approval-block`` class so the row's
+  // height stays stable when the real content swaps in (no shove of
+  // sibling rows mid-mouse-movement). Carries an aria-label so a
+  // screen reader announces the pending approval; the actual
+  // assertive ``announceApproval`` call in handleChildState fires
+  // once on the rising edge so the operator hears the demand even
+  // before the bulk fetch lands.
+  function renderApprovalPlaceholder(child) {
+    const block = document.createElement("div");
+    block.className = "approval-block approval-block-loading";
+    block.setAttribute("role", "region");
+    block.setAttribute(
+      "aria-label",
+      "Approval required for " +
+        (child.name || child.ws_id || "child") +
+        " — loading details",
+    );
+    const header = document.createElement("div");
+    header.className = "approval-header";
+    const pill = document.createElement("span");
+    pill.className = "approval-pill approval-pill-pending";
+    const spin = document.createElement("span");
+    spin.className = "approval-loading-spin";
+    pill.appendChild(spin);
+    pill.appendChild(document.createTextNode(" loading…"));
+    header.appendChild(pill);
+    block.appendChild(header);
+    return block;
+  }
+
   function renderApprovalBlock(child, detail) {
     if (!detail || !Array.isArray(detail.items) || detail.items.length === 0) {
       return null;
@@ -2562,6 +2594,16 @@
 
     const block = document.createElement("div");
     block.className = "approval-block";
+    block.setAttribute("role", "region");
+    block.setAttribute(
+      "aria-label",
+      "Approval required for " +
+        (child.name || child.ws_id || "child") +
+        " — " +
+        (primary && primary.header
+          ? primary.header
+          : items.length + " tool calls"),
+    );
 
     // Header line: pill + tool name(s) + tier:model
     const header = document.createElement("div");
@@ -2603,6 +2645,22 @@
       const conf = verdict.confidence;
       const confStr = typeof conf === "number" ? " " + conf.toFixed(2) : "";
       pill.textContent = (verdict.risk_level || "").toUpperCase() + confStr;
+      // SR-friendly label: spell out the risk level + optional
+      // confidence so a screen reader doesn't read "LOW 0.85" as
+      // a string of letters and a number. Visual text stays compact.
+      const riskWord =
+        riskCls === "crit"
+          ? "critical"
+          : riskCls === "high"
+            ? "high"
+            : riskCls === "med"
+              ? "medium"
+              : "low";
+      const confLabel =
+        typeof conf === "number"
+          ? ", confidence " + Math.round(conf * 100) + "%"
+          : "";
+      pill.setAttribute("aria-label", riskWord + " risk" + confLabel);
     }
     header.appendChild(pill);
 
@@ -2803,17 +2861,33 @@
         // Stale call_id \u2014 server has rolled to a new round, or
         // (more commonly) the approval was already resolved on
         // another channel and this click raced. Keep both buttons
-        // disabled until the urgent refresh lands and re-renders
-        // the row: the row is about to be replaced wholesale, so
-        // the disabled DOM is dropped along with it. Re-enabling
-        // here was the bug \u2014 it opened a window where rapid clicks
-        // hit the same already-resolved approval, each producing a
-        // fresh 409, looping until the live-bulk eventually cleared
-        // the row. On the rare path where the urgent refresh fails
-        // entirely, the operator can hit the Refresh button on the
-        // children panel to force a full reload.
+        // disabled until the refresh lands and re-renders the row:
+        // the row is about to be replaced wholesale, so the disabled
+        // DOM is dropped along with it. Re-enabling here was the bug
+        // \u2014 it opened a window where rapid clicks hit the same
+        // already-resolved approval, each producing a fresh 409,
+        // looping until the live-bulk eventually cleared the row.
+        // On the rare path where the refresh fails entirely, the
+        // operator can hit the Refresh button on the children panel
+        // to force a full reload. invalidateLiveBadge clears the
+        // cached entry so the next scheduleLiveFetch falls through
+        // the TTL gate (no cached entry to compare against); the
+        // standard 250ms debounce batches with any other in-flight
+        // pending ids.
+        // Inline note so the operator's "did my click work?" question
+        // gets answered without a noisy toast. Stays in the row until
+        // the refresh replaces the whole approval block.
+        const block = denyBtn.closest(".approval-block");
+        if (block && !block.querySelector(".approval-stale-note")) {
+          const note = document.createElement("div");
+          note.className = "approval-stale-note";
+          note.setAttribute("role", "status");
+          note.textContent =
+            "\u21bb already resolved elsewhere \u2014 refreshing\u2026";
+          block.appendChild(note);
+        }
         invalidateLiveBadge(targetWsId);
-        scheduleLiveFetch(targetWsId, { urgent: true });
+        scheduleLiveFetch(targetWsId);
         // Quiet console-warn for diagnostics; no toast \u2014 the
         // disappearing buttons / fresh row IS the operator-facing
         // signal, and a toast on every rapid-click 409 would just
@@ -2826,14 +2900,30 @@
       }
       // Optimistic clear \u2014 the next child_ws_state event will arrive
       // shortly and trigger a real refresh, but clearing locally
-      // makes the buttons disappear immediately on click.
+      // makes the buttons disappear immediately on click. The
+      // ``sseUpdatedAt`` bump is load-bearing: ``flushLiveFetches``'s
+      // merge guard preserves cleared pending_approval / _detail
+      // against an in-flight bulk fetch only while
+      // ``now - prev.sseUpdatedAt < SSE_AUTHORITATIVE_MS`` \u2014 without
+      // bumping, a bulk fetch landing in the gap reverts the
+      // cleared state from the upstream cache and the approve/deny
+      // pill flickers back. (Caught by /review bug-4.)
       const cached = liveBadgeCache.get(targetWsId);
       if (cached && cached.live) {
         cached.live = Object.assign({}, cached.live, {
           pending_approval: false,
           pending_approval_detail: null,
         });
-        liveBadgeCache.set(targetWsId, cached);
+        cached.sseUpdatedAt = Date.now();
+        _liveBadgeCacheSet(targetWsId, cached);
+      }
+      // Also clear the activity_state mirror so the row's "⚑ approval"
+      // badge in .meta disappears immediately too — without this, the
+      // row shows the badge with no buttons for ~50-150ms until the
+      // child_ws_approval_resolved push or next state event lands.
+      const childState = childrenState.get(targetWsId);
+      if (childState && childState.activity_state === "approval") {
+        childState.activity_state = "";
       }
       renderChildren();
     } catch (e) {
@@ -2902,7 +2992,49 @@
     return _childObserver;
   }
 
+  // Focus preservation helpers shared by _renderChildrenNow and
+  // _updateChildRow. ``replaceChildren()`` / ``replaceWith()`` blow
+  // away the focused element silently — without restore, the operator
+  // gets bounced to <body> mid-Tab whenever any state event fires for
+  // a row in the children tree.
+  function _captureRowFocusKey(scopeEl) {
+    const active = document.activeElement;
+    if (!active || !scopeEl || !scopeEl.contains(active)) return null;
+    const row = active.closest(".ch-row");
+    if (!row || !row.dataset.wsId) return null;
+    return {
+      wsId: row.dataset.wsId,
+      marker: active.className || active.tagName,
+    };
+  }
+
+  function _restoreRowFocus(scopeEl, focusKey) {
+    if (!focusKey || !scopeEl) return;
+    const sel = '.ch-row[data-ws-id="' + cssEscape(focusKey.wsId) + '"]';
+    const row = scopeEl.matches(sel) ? scopeEl : scopeEl.querySelector(sel);
+    if (!row) return;
+    let target = null;
+    if (focusKey.marker) {
+      // CSS.escape can't safely round-trip a class list with spaces,
+      // so we walk focusables and string-compare. A future refactor
+      // could swap to a stable ``data-focus-key`` attribute on each
+      // focusable element to dodge class-string identity entirely
+      // (see /review bug-2 — currently low risk because all
+      // focusables in renderChildRow / renderApprovalBlock carry
+      // single-class names).
+      const candidates = row.querySelectorAll("button, [tabindex], a, summary");
+      for (const el of candidates) {
+        if ((el.className || el.tagName) === focusKey.marker) {
+          target = el;
+          break;
+        }
+      }
+    }
+    if (target) target.focus({ preventScroll: true });
+  }
+
   function _renderChildrenNow() {
+    const focusKey = _captureRowFocusKey(childrenTreeEl);
     childrenTreeEl.setAttribute("aria-busy", "false");
     const rows = Array.from(childrenState.values());
     // Sort: non-terminal states first, then by name.
@@ -2934,7 +3066,42 @@
         else visibleChildIds.add(r.ws_id); // fallback: treat all visible
       });
     }
-    childrenCountEl.textContent = rows.length ? "(" + rows.length + ")" : "";
+    // Sidebar count: total + pending-approval annotation. The
+    // pending count is maintained incrementally on cache mutations
+    // (see ``pendingApprovalIds`` near the cache definition) so this
+    // is O(1) per render rather than an O(N) walk over the cache.
+    const pending = pendingApprovalIds.size;
+    childrenCountEl.textContent = rows.length
+      ? "(" +
+        rows.length +
+        (pending > 0 ? " · " + pending + " pending" : "") +
+        ")"
+      : "";
+    _restoreRowFocus(childrenTreeEl, focusKey);
+  }
+
+  // Targeted single-row update — used by handlers that only touch
+  // one row's state (verdict landing, approval resolved). Avoids
+  // the tree-wide rebuild ``_renderChildrenNow`` does so a 200-child
+  // tree doesn't re-paint 199 unaffected rows on every verdict
+  // arrival. Falls back to the full render if the row isn't in the
+  // DOM yet (first-time render). Preserves keyboard focus across
+  // the row swap — the same invariant ``_renderChildrenNow`` keeps
+  // for tree-wide rebuilds (caught by /review bug-3).
+  function _updateChildRow(childId) {
+    const sel = '.ch-row[data-ws-id="' + cssEscape(childId) + '"]';
+    const row = childrenTreeEl.querySelector(sel);
+    const entry = childrenState.get(childId);
+    if (row && entry) {
+      const focusKey = _captureRowFocusKey(row);
+      const replacement = renderChildRow(entry);
+      row.replaceWith(replacement);
+      const obs = _getChildObserver();
+      if (obs) obs.observe(replacement);
+      _restoreRowFocus(replacement, focusKey);
+    } else {
+      renderChildren();
+    }
   }
 
   function renderTaskRow(task) {
@@ -3065,15 +3232,9 @@
   const LIVE_BADGE_BULK_FLUSH_MS = LIVE_BADGE_DEBOUNCE_MS;
   const pendingLiveIds = new Set();
   let liveBadgeFlushTimer = null;
-  // Urgent-flush coalesce flag — N urgent calls in the same JS tick
-  // would otherwise issue N single-id bulk fetches (bulk endpoint
-  // accepts up to LIVE_BADGE_BULK_CAP ids per request). queueMicrotask
-  // batches them into one request that drains pendingLiveIds.
-  let urgentFlushScheduled = false;
 
-  function scheduleLiveFetch(childWsId, opts) {
+  function scheduleLiveFetch(childWsId) {
     if (!childWsId) return;
-    const urgent = !!(opts && opts.urgent);
     // Skip terminal-state children entirely — their live block will
     // never change again; fetching just burns a round-trip and caches
     // a stale value.  Renderer already styles closed/deleted rows.
@@ -3093,40 +3254,16 @@
       // change on any child triggers a fresh fetch → retry storm for
       // users who'll never have permission mid-session.
       if (cached.permanent) return;
-      // Urgent fetches bypass the TTL — used when a child enters
-      // approval state and the row needs the rich pending_approval_detail
-      // payload (call_id, items, judge_verdict) to render inline buttons.
-      // Waiting for the next 5s TTL window would leave the operator
-      // staring at "⚑ approval" with no way to act.
-      if (!urgent && Date.now() - cached.fetched < LIVE_BADGE_TTL_MS) return;
+      // TTL gate — slower-moving fields (tokens, context_ratio) refresh
+      // on this cadence. Approval state is event-driven (intent_verdict
+      // / approval_resolved push, plus the bulk fetch on initial
+      // approval entry); callers wanting a forced re-fetch
+      // (e.g. 409 stale-call_id retry) call invalidateLiveBadge first
+      // so the cache is empty and this gate falls through.
+      if (Date.now() - cached.fetched < LIVE_BADGE_TTL_MS) return;
     }
     if (!WS_ID_RE.test(childWsId)) return;
     pendingLiveIds.add(childWsId);
-    // Urgent: cancel the pending debounce and schedule a flush on
-    // the next microtask so N urgent calls in the same tick coalesce
-    // into one bulk request. Without the microtask hop, each urgent
-    // caller would drain pendingLiveIds with a single id and fire a
-    // separate fetch — defeating the bulk endpoint that accepts up
-    // to LIVE_BADGE_BULK_CAP ids per request.
-    if (urgent) {
-      if (liveBadgeFlushTimer !== null) {
-        clearTimeout(liveBadgeFlushTimer);
-        liveBadgeFlushTimer = null;
-      }
-      if (!urgentFlushScheduled) {
-        urgentFlushScheduled = true;
-        const flush = () => {
-          urgentFlushScheduled = false;
-          flushLiveFetches();
-        };
-        if (typeof queueMicrotask === "function") {
-          queueMicrotask(flush);
-        } else {
-          setTimeout(flush, 0);
-        }
-      }
-      return;
-    }
     if (liveBadgeFlushTimer !== null) return;
     liveBadgeFlushTimer = setTimeout(() => {
       liveBadgeFlushTimer = null;
@@ -3181,7 +3318,7 @@
             pending_approval_detail: prev.live.pending_approval_detail,
           });
         }
-        liveBadgeCache.set(id, {
+        _liveBadgeCacheSet(id, {
           live: mergedLive,
           fetched: now,
           // Denied ids are permission/identity misses — mark permanent
@@ -3209,7 +3346,7 @@
       const now = Date.now();
       ids.forEach((id) => {
         const prev = liveBadgeCache.get(id);
-        liveBadgeCache.set(id, {
+        _liveBadgeCacheSet(id, {
           live: null,
           fetched: now,
           permanent: isPermanent,
@@ -3221,7 +3358,7 @@
   }
 
   function invalidateLiveBadge(childWsId) {
-    liveBadgeCache.delete(childWsId);
+    _liveBadgeCacheDelete(childWsId);
   }
 
   // --- SSE handlers for child_ws_* events ----------------------------
@@ -3261,44 +3398,54 @@
     if (ev.node_id) existing.node_id = ev.node_id;
     childrenState.set(childId, existing);
     _touchChild(childId);
-    // pending_approval_detail rides on the ws_state event when an
-    // approval is pending (see turnstone/server.py
-    // WebUI._broadcast_state — gated on ``_pending_approval is not
-    // None``, so absent on the steady state and possibly null on a
-    // node mid-rolling-upgrade).  When present we mutate
-    // liveBadgeCache directly here — inline approve/deny buttons
-    // render in lockstep with the activity_state transition without
-    // a separate live-bulk fetch.  The cache entry is tagged
-    // ``sseUpdatedAt`` so a bulk-poll landing within
-    // SSE_AUTHORITATIVE_MS preserves the SSE-supplied fields
-    // (the upstream /dashboard cache's ~2s TTL would otherwise
-    // clobber a fresh transition with pre-transition state).
-    const pendingApproval = existing.activity_state === "approval";
-    const evDetail =
-      ev.pending_approval_detail !== undefined
-        ? ev.pending_approval_detail
-        : null;
+    // Track ``pending_approval`` flag from BOTH state and
+    // activity_state. ``state="attention"`` is the canonical signal
+    // that the workstream needs operator attention; ``activity_state
+    // ="approval"`` is the secondary signal set by approve_tools.
+    // We trust either: in practice the worker thread can fire the
+    // state transition before approve_tools has updated activity_state
+    // (the two writes happen on different lines under different
+    // locks), so a state-attention event with empty activity_state
+    // is a real and common case — checking only activity_state
+    // misses 30+ children all sitting in attention. (Caught manual
+    // testing: 30 children all state=attention rendered with no
+    // approval blocks because pendingApproval was always false.)
+    const pendingApproval =
+      existing.state === "attention" || existing.activity_state === "approval";
     const cached = liveBadgeCache.get(childId);
-    // When pending, prefer SSE-supplied detail.  If SSE didn't
-    // carry it (rare race; e.g. a node mid-rolling-upgrade), keep
-    // any existing cached detail rather than blanking the row —
-    // the next bulk-poll catches up.  When not pending, hard-clear.
-    let nextDetail;
-    if (pendingApproval) {
-      nextDetail =
-        evDetail !== null
-          ? evDetail
-          : cached && cached.live
-            ? cached.live.pending_approval_detail
-            : null;
-    } else {
-      nextDetail = null;
+    const cachedLive = (cached && cached.live) || {};
+    // Rising-edge detection BEFORE we mutate the cache. The chat-pane
+    // tool batches already announce assertively
+    // (renderApprovalDock / appendToolBatch); the children-tree was
+    // silent for SR users — fixing that here so a blind operator
+    // hears the demand for action.
+    const wasPendingApproval = cachedLive.pending_approval === true;
+    if (pendingApproval && !wasPendingApproval) {
+      _announceAssertive(
+        "Approval required: " + (existing.name || childId.slice(0, 8)),
+      );
     }
-    const nextLive = Object.assign({}, (cached && cached.live) || {}, {
+    const nextLive = Object.assign({}, cachedLive, {
       pending_approval: pendingApproval,
-      pending_approval_detail: nextDetail,
     });
-    liveBadgeCache.set(childId, {
+    // Off-approval transition is the ONLY case here that
+    // authoritatively writes a value the bulk fetch must not
+    // resurrect (a stale bulk-fetch landing within
+    // SSE_AUTHORITATIVE_MS would otherwise re-render the cleared
+    // approval block). Setting ``pending_approval=true`` does NOT
+    // claim cache authority — the bulk fetch is the source of
+    // truth for ``pending_approval_detail``, and bumping
+    // sseUpdatedAt here makes the merge guard in flushLiveFetches
+    // preserve our stale (often null) detail over the bulk
+    // fetch's actual data, leaving the row stuck on the loading
+    // placeholder. (Caught when the screenshot showed buttons
+    // briefly then loading replaced them.)
+    const detailClearedAuthoritatively =
+      !pendingApproval && cachedLive.pending_approval === true;
+    if (!pendingApproval) {
+      nextLive.pending_approval_detail = null;
+    }
+    _liveBadgeCacheSet(childId, {
       live: nextLive,
       // Preserve prior bulk-poll fetched timestamp so a fresh SSE
       // tick doesn't artificially extend the 5s TTL gate in
@@ -3306,7 +3453,11 @@
       // moving fields (tokens, context_ratio) on its own schedule.
       fetched: cached ? cached.fetched : 0,
       permanent: !!(cached && cached.permanent),
-      sseUpdatedAt: Date.now(),
+      sseUpdatedAt: detailClearedAuthoritatively
+        ? Date.now()
+        : cached
+          ? cached.sseUpdatedAt || 0
+          : 0,
     });
     renderChildren();
     // Do NOT invalidateLiveBadge on routine state ticks — that
@@ -3344,6 +3495,104 @@
     renderChildren();
   }
 
+  // Stage 3 Step 7 — verdict + approval reducer.
+  //
+  // Both write directly to liveBadgeCache and skip scheduleLiveFetch,
+  // which sidesteps the visibility gate in that path so off-screen
+  // rows pick up the new cache value the moment they scroll back in.
+  // Both are idempotent — the bulk-fetch (single source of truth for
+  // the items list) and the explicit intent_verdict event can deliver
+  // overlapping data; receiving twice re-stamps the same value.
+  //
+  // Both call ``_updateChildRow`` (targeted single-row swap) instead
+  // of ``renderChildren`` so a verdict landing on row 3 doesn't rebuild
+  // every other row in a 200-child sidebar — which would also blow
+  // away keyboard focus on whatever row the operator was tabbing to.
+
+  function handleChildIntentVerdict(ev) {
+    const childId = ev.child_ws_id || ev.ws_id;
+    if (!childId) return;
+    const verdict = ev.verdict || {};
+    const callId = verdict.call_id || "";
+    if (!callId) return;
+    const cached = liveBadgeCache.get(childId);
+    const cachedLive = (cached && cached.live) || {};
+    const detail = cachedLive.pending_approval_detail;
+    if (!detail) {
+      // No pending_approval_detail to stamp the verdict onto. The
+      // verdict is still durable in storage; the next bulk fetch
+      // will hydrate the detail and include the verdict via the
+      // existing serialize path.
+      return;
+    }
+    // Stamp on the matching item (UI render reads judge_verdict per
+    // item) AND on the by-call_id map (matches the
+    // serialize_pending_approval_detail shape).
+    const items = Array.isArray(detail.items) ? detail.items : [];
+    for (const item of items) {
+      if (item && item.call_id === callId) {
+        item.judge_verdict = verdict;
+        break;
+      }
+    }
+    if (!detail.llm_verdicts) detail.llm_verdicts = {};
+    detail.llm_verdicts[callId] = verdict;
+    // judge_pending flips false once every item has a verdict —
+    // matches the server-side serializer's logic.
+    if (items.length > 0) {
+      detail.judge_pending = !items.every((it) => it && it.judge_verdict);
+    }
+    _liveBadgeCacheSet(childId, {
+      live: cachedLive,
+      fetched: cached ? cached.fetched : 0,
+      permanent: !!(cached && cached.permanent),
+      sseUpdatedAt: Date.now(),
+    });
+    _updateChildRow(childId);
+  }
+
+  function handleChildApprovalResolved(ev) {
+    const childId = ev.child_ws_id || ev.ws_id;
+    if (!childId) return;
+    const cached = liveBadgeCache.get(childId);
+    const cachedLive = (cached && cached.live) || {};
+    cachedLive.pending_approval = false;
+    cachedLive.pending_approval_detail = null;
+    _liveBadgeCacheSet(childId, {
+      live: cachedLive,
+      fetched: cached ? cached.fetched : 0,
+      permanent: !!(cached && cached.permanent),
+      sseUpdatedAt: Date.now(),
+    });
+    _updateChildRow(childId);
+  }
+
+  // Push path for the initial approval items — eliminates the
+  // bulk-fetch race that previously left rows stuck on a loading
+  // placeholder when the bulk fetch landed in the gap between the
+  // state transition to ATTENTION and ``_pending_approval`` being
+  // set inside ``approve_tools``. Stamps the items into the cache
+  // directly; the bulk fetch remains as a reconnect / refresh
+  // fallback. Idempotent — receiving the same approve_request twice
+  // re-stamps the same detail.
+  function handleChildApproveRequest(ev) {
+    const childId = ev.child_ws_id || ev.ws_id;
+    if (!childId) return;
+    const detail = ev.detail || null;
+    if (!detail) return;
+    const cached = liveBadgeCache.get(childId);
+    const cachedLive = (cached && cached.live) || {};
+    cachedLive.pending_approval = true;
+    cachedLive.pending_approval_detail = detail;
+    _liveBadgeCacheSet(childId, {
+      live: cachedLive,
+      fetched: cached ? cached.fetched : 0,
+      permanent: !!(cached && cached.permanent),
+      sseUpdatedAt: Date.now(),
+    });
+    _updateChildRow(childId);
+  }
+
   // Periodic sweep of stale terminal rows.  Operator tabs left open all
   // day would otherwise accumulate entries for every child the
   // coordinator ever spawned — rows the user can still see (state !=
@@ -3359,7 +3608,7 @@
       if (terminal && now - lastSeen > CHILDREN_TERMINAL_GRACE_MS) {
         childrenState.delete(id);
         childrenLastSeen.delete(id);
-        liveBadgeCache.delete(id);
+        _liveBadgeCacheDelete(id);
         visibleChildIds.delete(id);
         removed += 1;
       }
@@ -3375,7 +3624,7 @@
         const id = byAge[i][0];
         childrenState.delete(id);
         childrenLastSeen.delete(id);
-        liveBadgeCache.delete(id);
+        _liveBadgeCacheDelete(id);
         visibleChildIds.delete(id);
         removed += 1;
       }
@@ -3388,7 +3637,7 @@
 
   if (childrenRefreshBtn) {
     childrenRefreshBtn.addEventListener("click", () => {
-      liveBadgeCache.clear();
+      _liveBadgeCacheClear();
       // Explicit refresh wipes SSE-discovered rows the server no
       // longer knows about — the operator asked for a clean snapshot.
       loadChildren({ replace: true });

turnstone/console/static/coordinator/index.html

@@ -181,6 +181,42 @@
     font-family: var(--font-mono);
     font-size: 11px;
     line-height: 1.4;
+    /* Smooth the auto-expand on judge-verdict landing so the eye can
+       track sibling rows shifting down rather than them snapping. */
+    transition: max-height 200ms ease-out;
+    max-height: 80vh;
+    overflow: hidden;
+  }
+  /* Loading-state placeholder rendered while the bulk fetch is
+     in-flight. Same outer container as the real block so the row
+     height is stable when the real content swaps in. */
+  .ch-row .approval-block-loading {
+    color: var(--ink-3);
+  }
+  .ch-row .approval-loading-spin {
+    display: inline-block;
+    width: 8px;
+    height: 8px;
+    border-radius: 50%;
+    border: 1.5px solid var(--accent);
+    border-top-color: transparent;
+    margin-right: 5px;
+    animation: ts-spin 0.9s linear infinite;
+    vertical-align: middle;
+  }
+  /* Inline status note shown when the operator clicks Approve/Deny
+     but the call_id is stale (already resolved on another channel).
+     Quieter than a toast — the row is about to be replaced wholesale
+     by the refresh, so this only needs to bridge ~350ms. */
+  .ch-row .approval-stale-note {
+    font-size: 11px;
+    color: var(--ink-3);
+    font-style: italic;
+    margin-top: 4px;
+  }
+  @media (prefers-reduced-motion: reduce) {
+    .ch-row .approval-block { transition: none; }
+    .ch-row .approval-loading-spin { animation: none; }
   }
   /* Auto-approved pill — same 22px indent as .approval-block / .meta
      so a child row showing both stacks cleanly.  Lower visual weight
@@ -442,6 +478,22 @@
      link feed.css (no .feed-item usage). */
   @keyframes ts-spin { to { transform: rotate(360deg); } }
 
+  /* SSE queue depth indicator — populated by the periodic
+     ``_queue_stats`` event in the events SSE handler. Color escalates
+     so a backed-up queue (slow consumer / browser throttling
+     background tabs) is glanceable from the status bar. */
+  .ws-status-bar .ws-sb-sse-queue {
+    color: var(--ink-3);
+    font-size: 11px;
+  }
+  .ws-status-bar .ws-sb-sse-queue.warn {
+    color: var(--warn);
+  }
+  .ws-status-bar .ws-sb-sse-queue.danger {
+    color: var(--err);
+    font-weight: 600;
+  }
+
   /* Sidebar mobile toggle (desktop hides; mobile shows via media query
      below). */
   #coord-sidebar-toggle {
@@ -542,6 +594,12 @@
       <span id="coord-sb-tokens" class="ws-sb-tokens" aria-label="Token usage">0 / —</span>
       <span id="coord-sb-tools" class="ws-sb-tools" aria-label="Tool calls this turn">0 tools</span>
       <span id="coord-sb-turns" class="ws-sb-turns" aria-label="Conversation turn">turn 0</span>
+      <!-- Server-side per-coord SSE listener queue depth — populated
+           by the periodic ``_queue_stats`` event. Hidden until the
+           first sample lands so the bar isn't permanently cluttered
+           on connections that never report depth. -->
+      <span id="coord-sb-sse-queue" class="ws-sb-sse-queue"
+            aria-label="SSE queue depth" hidden>queue 0/500</span>
     </div>
 
     <!-- Composer DOM is built by shared_static/composer.js into this mount. -->

turnstone/core/child_source.py

@@ -0,0 +1,264 @@
+"""Strategy interface for delivering child workstream lifecycle events.
+
+Two implementations bind to the unified :class:`ChildrenRegistry`:
+
+- :class:`SameNodeChildSource` — subscribes to a :class:`SessionManager`'s
+  state-change callbacks. In-process, no transport. For interactive
+  workstreams that spawn children locally (no cluster routing).
+- :class:`ClusterChildSource` — subscribes to a :class:`ClusterCollector`'s
+  listener channel and runs a daemon thread that drains the queue and
+  pushes events to the sink. For coordinator workstreams whose
+  children are routed across the cluster by hash bucket.
+
+The strategy doesn't translate events into UI-shaped payloads; that's
+the sink's job. This split keeps the strategy generic across kinds and
+lets the consumer (e.g. ``CoordinatorAdapter._dispatch_child_event``)
+own the per-kind translation.
+
+Sink signature: ``Callable[[dict[str, Any]], None]``. The sink is
+responsible for filtering by registry membership; strategies push raw
+events without registry-side filtering so the sink can decide whether
+to act based on its own state.
+"""
+
+from __future__ import annotations
+
+import queue
+import threading
+from typing import TYPE_CHECKING, Any, Protocol
+
+from turnstone.core.log import get_logger
+
+if TYPE_CHECKING:
+    from collections.abc import Callable, Iterable
+
+    from turnstone.core.children_registry import ChildrenRegistry
+    from turnstone.core.workstream import WorkstreamState
+
+log = get_logger(__name__)
+
+
+class ChildSource(Protocol):
+    """Subscription strategy for child workstream lifecycle events."""
+
+    def start(self, sink: Callable[[dict[str, Any]], None]) -> None:
+        """Begin delivering events to ``sink``. Idempotent."""
+
+    def shutdown(self) -> None:
+        """Stop the strategy. Idempotent; safe to call multiple times."""
+
+
+class _CollectorProtocol(Protocol):
+    """Subset of :class:`ClusterCollector` that :class:`ClusterChildSource` consumes.
+
+    Defined here (not imported) to keep ``turnstone/core/`` free of
+    ``turnstone/console/`` imports AND to let test fakes satisfy the
+    type signature without subclassing the real collector.
+    """
+
+    def get_snapshot_and_register(self, q: queue.Queue[dict[str, Any]]) -> dict[str, Any]: ...
+
+    def unregister_listener(self, q: queue.Queue[dict[str, Any]]) -> None: ...
+
+
+class _ManagerProtocol(Protocol):
+    """Subset of :class:`SessionManager` that :class:`SameNodeChildSource` consumes.
+
+    Lets test fakes participate in the strategy's typed surface
+    without forcing a full SessionManager construction.
+    """
+
+    def subscribe_to_state(self, callback: Callable[[str, WorkstreamState], None]) -> None: ...
+
+    def unsubscribe_from_state(self, callback: Callable[[str, WorkstreamState], None]) -> None: ...
+
+
+class SameNodeChildSource:
+    """In-process child events via :class:`SessionManager` state observer.
+
+    Subscribes to the manager's state-change callbacks (registered via
+    :meth:`SessionManager.subscribe_to_state`). For each transition on
+    a workstream that's a known child (per the
+    :class:`ChildrenRegistry` reverse index), synthesises a
+    cluster-state-shaped event and pushes it to the sink.
+
+    Used by interactive workstreams when they gain spawn capability —
+    children live on the same node as the parent, so no cluster routing
+    is needed and event fan-out is in-process.
+    """
+
+    def __init__(
+        self,
+        manager: _ManagerProtocol,
+        registry: ChildrenRegistry,
+    ) -> None:
+        self._manager = manager
+        self._registry = registry
+        self._sink: Callable[[dict[str, Any]], None] | None = None
+        self._callback: Callable[[str, WorkstreamState], None] | None = None
+
+    def start(self, sink: Callable[[dict[str, Any]], None]) -> None:
+        if self._callback is not None:
+            return  # idempotent — already started
+        self._sink = sink
+
+        def _on_state(ws_id: str, state: WorkstreamState) -> None:
+            sink_fn = self._sink
+            if sink_fn is None:
+                return
+            # Cheap pre-filter: skip dispatch for transitions on
+            # workstreams that aren't children of any in-memory parent.
+            # ``has_children`` is a lock-free dict-truthiness read; the
+            # ``parent_for`` call below would otherwise acquire the
+            # registry lock on every state change even when no
+            # children exist (the steady state for an interactive
+            # manager). The cluster strategy can't pre-filter because
+            # the collector queue carries all events; here we have the
+            # information to skip the synthesis entirely.
+            if not self._registry.has_children():
+                return
+            if self._registry.parent_for(ws_id) is None:
+                return
+            # ``pending_approval_detail`` deliberately omitted — the
+            # field was removed from cluster_state end-to-end in the
+            # Stage 3 cleanup pass. Approval items arrive via bulk
+            # fetch; verdicts via the explicit intent_verdict event;
+            # resolution via approval_resolved.
+            event = {
+                "type": "cluster_state",
+                "ws_id": ws_id,
+                "state": state.value,
+                "node_id": "",
+                "tokens": 0,
+                "activity_state": "",
+            }
+            try:
+                sink_fn(event)
+            except Exception:
+                log.debug("same_node_child_source.sink_failed", exc_info=True)
+
+        self._callback = _on_state
+        self._manager.subscribe_to_state(_on_state)
+
+    def shutdown(self) -> None:
+        cb = self._callback
+        if cb is None:
+            return
+        try:
+            self._manager.unsubscribe_from_state(cb)
+        except Exception:
+            log.debug("same_node_child_source.unsubscribe_failed", exc_info=True)
+        self._callback = None
+        self._sink = None
+
+
+class ClusterChildSource:
+    """Cross-node child events via :class:`ClusterCollector` subscription.
+
+    Refactor of the existing fan-out machinery from
+    ``CoordinatorAdapter`` (was ``_collector_queue`` +
+    ``_fanout_thread`` + ``_fanout_loop``). Subscribes as a listener on
+    the collector's broadcast channel and runs a daemon thread that
+    drains the queue, pushing each event to the sink.
+
+    On :meth:`start`, also primes the registry from the collector's
+    snapshot so a parent that re-installs after a console restart sees
+    its already-live children without waiting for the next state tick.
+    The ``parents_provider`` callback returns the set of in-memory
+    parent ws_ids for snapshot filtering — only children whose parent
+    is currently installed get merged.
+    """
+
+    def __init__(
+        self,
+        collector: _CollectorProtocol,
+        registry: ChildrenRegistry,
+        *,
+        parents_provider: Callable[[], Iterable[str]],
+    ) -> None:
+        self._collector = collector
+        self._registry = registry
+        self._parents_provider = parents_provider
+        self._sink: Callable[[dict[str, Any]], None] | None = None
+        self._queue: queue.Queue[dict[str, Any]] | None = None
+        self._thread: threading.Thread | None = None
+        self._stop = threading.Event()
+
+    def start(self, sink: Callable[[dict[str, Any]], None]) -> None:
+        if self._thread is not None and self._thread.is_alive():
+            return  # idempotent — already started
+        self._sink = sink
+        self._queue = queue.Queue(maxsize=1000)
+        snapshot = self._collector.get_snapshot_and_register(self._queue)
+        self._prime_from_snapshot(snapshot)
+        self._stop.clear()
+        t = threading.Thread(
+            target=self._loop,
+            name="cluster-child-source",
+            daemon=True,
+        )
+        self._thread = t
+        t.start()
+
+    def shutdown(self) -> None:
+        self._stop.set()
+        t = self._thread
+        q = self._queue
+        coll = self._collector
+        self._thread = None
+        self._queue = None
+        if coll is not None and q is not None:
+            try:
+                coll.unregister_listener(q)
+            except Exception:
+                log.debug(
+                    "cluster_child_source.unregister_listener_failed",
+                    exc_info=True,
+                )
+        if t is not None:
+            t.join(timeout=2.0)
+        self._sink = None
+
+    def _loop(self) -> None:
+        q = self._queue
+        if q is None:
+            return
+        while not self._stop.is_set():
+            try:
+                event = q.get(timeout=1.0)
+            except queue.Empty:
+                continue
+            sink = self._sink
+            if sink is None:
+                continue
+            try:
+                sink(event)
+            except Exception:
+                log.debug("cluster_child_source.dispatch_failed", exc_info=True)
+
+    def _prime_from_snapshot(self, snapshot: dict[str, Any]) -> None:
+        """Populate the registry from a collector snapshot.
+
+        For every workstream in the snapshot whose ``parent_ws_id``
+        names a currently-installed parent (per ``parents_provider``),
+        merge it into the registry. Caller-installed parents that
+        appear in the snapshot are seeded; unknown parents are skipped
+        — they'll be picked up by the live fan-out path once their
+        ``ws_created`` event arrives.
+        """
+        nodes = snapshot.get("nodes", []) if isinstance(snapshot, dict) else []
+        if not nodes:
+            return
+        known_parents = set(self._parents_provider())
+        if not known_parents:
+            return
+        by_parent: dict[str, list[str]] = {}
+        for node in nodes:
+            for entry in node.get("workstreams", []) or []:
+                parent = entry.get("parent_ws_id") or ""
+                child_id = entry.get("id") or ""
+                if not parent or not child_id or parent not in known_parents:
+                    continue
+                by_parent.setdefault(parent, []).append(child_id)
+        for parent, kids in by_parent.items():
+            self._registry.merge_children(parent, kids)

turnstone/core/children_registry.py

@@ -0,0 +1,177 @@
+"""Universal parent → children registry for SessionManager.
+
+Pure data + lookups; no IO, no transport. Lifted from
+:class:`turnstone.console.coordinator_adapter.CoordinatorAdapter` where
+it lived bound to the coordinator kind. The lift is what lets the
+``ChildSource`` strategies (Step 2) plug into a single shared primitive
+regardless of whether children are local (interactive) or cluster-routed
+(coordinator).
+
+Storage rebuild and snapshot priming happen in the caller — typically
+the ``ChildSource`` implementation that owns the relevant transport.
+The registry exposes :meth:`merge_children` for bulk seeding so callers
+that compute child id lists from any source can feed them in without
+the registry needing to know about storage shapes or collector
+snapshots.
+
+Threading: every public method is internally locked. Helpers suffixed
+``_locked`` require the caller to already hold :attr:`_lock`.
+"""
+
+from __future__ import annotations
+
+import threading
+from typing import TYPE_CHECKING, Any
+
+if TYPE_CHECKING:
+    from collections.abc import Iterable
+
+
+class ChildrenRegistry:
+    """Tracks parent → children + reverse lookup for in-memory parents.
+
+    The forward index (``_children``) is parent_ws_id → set of child ws_ids.
+    The reverse index (``_child_to_parent``) is child_ws_id → parent_ws_id.
+    The presence map (``_active``) is parent_ws_id → UI ref, used by the
+    dispatch path to atomically check-and-route in one lock acquisition.
+
+    Closed / deleted children stay in the registry until their owning
+    parent is uninstalled — the tree UI keeps rendering them grayed out;
+    state authority lives in storage, not here.
+    """
+
+    def __init__(self) -> None:
+        self._children: dict[str, set[str]] = {}
+        self._child_to_parent: dict[str, str] = {}
+        self._lock = threading.Lock()
+        self._active: dict[str, Any] = {}
+
+    # ------------------------------------------------------------------
+    # Lifecycle — install / uninstall a parent
+    # ------------------------------------------------------------------
+
+    def install(self, parent_ws_id: str, ui: Any) -> None:
+        """Seed the forward set + presence map for a new parent.
+
+        Idempotent — re-installing re-points the UI but leaves the
+        existing child set intact. Mirrors the original
+        ``_install_coord_registry`` semantics so a coordinator that
+        rehydrates after a crash doesn't lose its known-children.
+        """
+        with self._lock:
+            self._children.setdefault(parent_ws_id, set())
+            self._active[parent_ws_id] = ui
+
+    def uninstall(self, parent_ws_id: str) -> None:
+        """Drop a parent: forward set, reverse-index entries, presence.
+
+        No-op if the parent is unknown. Used by close / eviction paths.
+        """
+        with self._lock:
+            self._uninstall_locked(parent_ws_id)
+
+    # ------------------------------------------------------------------
+    # Mutation — register children under a parent
+    # ------------------------------------------------------------------
+
+    def add_child(self, parent_ws_id: str, child_ws_id: str) -> Any | None:
+        """Register a child under a parent. Returns parent's UI or None.
+
+        Returns the parent's UI on success (so the dispatch path can
+        atomically check-and-route in one lock acquisition). Returns
+        ``None`` if the parent isn't installed (concurrent close /
+        eviction) or if the child is already registered (duplicate
+        ws_created from the cluster fan-out).
+        """
+        with self._lock:
+            ui = self._active.get(parent_ws_id)
+            if ui is None:
+                return None
+            existing = self._children.setdefault(parent_ws_id, set())
+            if child_ws_id in existing:
+                return None
+            existing.add(child_ws_id)
+            self._child_to_parent[child_ws_id] = parent_ws_id
+            return ui
+
+    def merge_children(self, parent_ws_id: str, child_ws_ids: Iterable[str]) -> None:
+        """Bulk-merge child_ids under a parent. Idempotent.
+
+        Sole bulk write-path — used by both storage-seeded rebuilds and
+        snapshot-seeded priming so reverse-index ordering invariants
+        hold regardless of which seed source races first.
+        """
+        with self._lock:
+            self._merge_locked(parent_ws_id, child_ws_ids)
+
+    # ------------------------------------------------------------------
+    # Lookups
+    # ------------------------------------------------------------------
+
+    def parent_for(self, child_ws_id: str) -> str | None:
+        """Reverse lookup: which parent owns this child? O(1)."""
+        with self._lock:
+            return self._child_to_parent.get(child_ws_id)
+
+    def has_children(self) -> bool:
+        """Lock-free fast path: any child registered under any parent?
+
+        Reads ``bool(self._child_to_parent)`` without taking the lock.
+        Dict-truthiness is a single GIL-atomic read, so callers on
+        the hot state-broadcast path can short-circuit without paying
+        the lock acquisition when the registry is empty (the steady
+        state for an interactive manager today). The answer is best-
+        effort — if a child is added concurrently with the read the
+        caller may falsely return ``False``, but the next state event
+        will pick up the change correctly.
+        """
+        return bool(self._child_to_parent)
+
+    def children_of(self, parent_ws_id: str) -> list[str]:
+        """Snapshot copy of the parent's child ws_ids.
+
+        Returned list is a copy so callers can iterate without holding
+        the registry lock during per-child work. A mutation racing with
+        the snapshot either lands before (included) or after (excluded)
+        — both outcomes are safe for cascade-style dispatch.
+        """
+        with self._lock:
+            child_set = self._children.get(parent_ws_id)
+            return list(child_set) if child_set else []
+
+    def ui_for(self, parent_ws_id: str) -> Any | None:
+        """Look up the UI registered for a parent."""
+        with self._lock:
+            return self._active.get(parent_ws_id)
+
+    def parents(self) -> list[str]:
+        """Snapshot copy of installed parent ws_ids."""
+        with self._lock:
+            return list(self._active)
+
+    # ------------------------------------------------------------------
+    # Locked helpers — caller must hold ``self._lock``
+    # ------------------------------------------------------------------
+
+    def _merge_locked(self, parent_ws_id: str, child_ws_ids: Iterable[str]) -> None:
+        """Idempotent merge under caller's lock. Empty/falsy ids skipped."""
+        existing = self._children.setdefault(parent_ws_id, set())
+        for cid in child_ws_ids:
+            if cid and cid not in existing:
+                existing.add(cid)
+                self._child_to_parent[cid] = parent_ws_id
+
+    def _uninstall_locked(self, parent_ws_id: str) -> None:
+        """Pop forward set + presence + own reverse-index entries.
+
+        Defensive: only clears reverse entries that still point at
+        ``parent_ws_id``. Schema-shaped reassignments (rare but
+        possible) shouldn't orphan the new owner's entry.
+        """
+        child_set = self._children.pop(parent_ws_id, None)
+        self._active.pop(parent_ws_id, None)
+        if child_set is None:
+            return
+        for cid in child_set:
+            if self._child_to_parent.get(cid) == parent_ws_id:
+                self._child_to_parent.pop(cid, None)

turnstone/core/session_manager.py

@@ -214,13 +214,21 @@ class SessionManager:
         # manager never reads them.
         self._active_id: str | None = None
         self._eviction_count: int = 0
-        # Optional state-change observer. The CLI sets this to a
-        # callback that prints a background-attention notification
-        # when a non-focused workstream transitions to ATTENTION.
-        # Web/coord paths use the event_emitter's emit_state for their
-        # own fan-out; this is a second, manager-level hook for callers
-        # that don't consume SSE.
-        self._on_state_change: Callable[[str, WorkstreamState], None] | None = None
+        # State-change subscribers. Multi-subscriber to support the
+        # CLI's background-attention notification AND the in-process
+        # ``SameNodeChildSource`` strategy that delivers child
+        # workstream state changes to a parent's UI without going
+        # through the cluster bus. Each callback fires under
+        # exception-suppression so one failing subscriber doesn't
+        # block the others. Subscribers register via
+        # :meth:`subscribe_to_state`. ``_state_subscribers_lock``
+        # guards mutation + snapshot — set_state copies the list
+        # under the lock then iterates the snapshot unlocked so a
+        # slow subscriber doesn't block subscribe/unsubscribe (and
+        # so concurrent subscribe/unsubscribe during a state event
+        # can't shift the iterator's index — caught by /review bug-1).
+        self._state_subscribers: list[Callable[[str, WorkstreamState], None]] = []
+        self._state_subscribers_lock = threading.Lock()
 
     # ------------------------------------------------------------------
     # Properties
@@ -816,9 +824,36 @@ class SessionManager:
                     log.debug("session_mgr.state_update_failed ws=%s", ws_id[:8], exc_info=True)
         if self._event_emitter is not None:
             self._event_emitter.emit_state(ws, state)
-        if self._on_state_change is not None:
+        # Snapshot under the subscribers lock so concurrent
+        # subscribe / unsubscribe can't shift the iterator's index
+        # mid-dispatch (skipping or repeating callbacks). Iterate
+        # the snapshot WITHOUT the lock so a slow callback doesn't
+        # block subscribe / unsubscribe.
+        with self._state_subscribers_lock:
+            subscribers = list(self._state_subscribers)
+        for callback in subscribers:
             with contextlib.suppress(Exception):
-                self._on_state_change(ws_id, state)
+                callback(ws_id, state)
+
+    # ------------------------------------------------------------------
+    # State-change subscription
+    # ------------------------------------------------------------------
+
+    def subscribe_to_state(self, callback: Callable[[str, WorkstreamState], None]) -> None:
+        """Register ``callback`` to fire on every workstream state change.
+
+        Multiple subscribers are supported and fire in registration order.
+        Each callback is wrapped in exception-suppression so a failing
+        subscriber doesn't block the others. Use
+        :meth:`unsubscribe_from_state` to remove.
+        """
+        with self._state_subscribers_lock:
+            self._state_subscribers.append(callback)
+
+    def unsubscribe_from_state(self, callback: Callable[[str, WorkstreamState], None]) -> None:
+        """Remove a previously-registered state-change callback. No-op if absent."""
+        with self._state_subscribers_lock, contextlib.suppress(ValueError):
+            self._state_subscribers.remove(callback)
 
     def cancel(self, ws_id: str) -> bool:
         """Cancel in-flight generation and unblock any pending approval / plan.

turnstone/core/session_routes.py

@@ -1377,9 +1377,14 @@ def make_events_handler(cfg: SessionEndpointConfig) -> Handler:
         import asyncio
         import json
         import queue
+        import time
 
         from sse_starlette import EventSourceResponse
 
+        from turnstone.core.session_ui_base import (
+            _DEFAULT_LISTENER_QUEUE_MAX,
+        )
+
         if cfg.permission_gate is not None:
             err = cfg.permission_gate(request)
             if err is not None:
@@ -1474,9 +1479,32 @@ def make_events_handler(cfg: SessionEndpointConfig) -> Handler:
                 # cancel-detection latency the timeout would otherwise
                 # gate; shortening to 1s 5x'd the wakeup rate without
                 # any client-observable benefit).
+                #
+                # Periodic ``_queue_stats`` emission carries the
+                # per-tab listener queue depth so the browser can
+                # surface backed-up SSE consumers in the status bar
+                # (slow tab / browser throttling / network stall
+                # would otherwise be invisible until events visibly
+                # stop arriving). 2s cadence is observational — the
+                # JSON payload is tiny and won't measurably grow the
+                # queue itself.
+                queue_stats_interval_s = 2.0
+                last_queue_stats_at = 0.0
                 while True:
                     if await request.is_disconnected():
                         return
+                    now_mono = time.monotonic()
+                    if now_mono - last_queue_stats_at >= queue_stats_interval_s:
+                        last_queue_stats_at = now_mono
+                        yield {
+                            "data": json.dumps(
+                                {
+                                    "type": "_queue_stats",
+                                    "depth": client_queue.qsize(),
+                                    "max": _DEFAULT_LISTENER_QUEUE_MAX,
+                                }
+                            )
+                        }
                     try:
                         event = await loop.run_in_executor(
                             live_executor,

turnstone/core/session_ui_base.py

@@ -41,6 +41,56 @@ log = get_logger(__name__)
 # from bloating memory.
 _DEFAULT_LISTENER_QUEUE_MAX = 500
 
+# Event types whose delivery is load-bearing for correctness — not
+# observational fluff. When a listener queue is full, ``_enqueue``
+# evicts the oldest event to make room for these rather than dropping
+# them on the floor (the default ``queue.Full`` swallow). Token-stream
+# / status / activity events drop equally as today; an approval
+# request or verdict landing during a chatty mid-generation burst
+# would otherwise be silently lost on a backed-up tab.
+_CRITICAL_EVENT_TYPES = frozenset(
+    {
+        "intent_verdict",
+        "intent_pending",
+        "approval_resolved",
+        "approve_request",
+        "plan_review",
+        "ws_closed",
+        "child_ws_intent_verdict",
+        "child_ws_approval_resolved",
+        "child_ws_closed",
+        "child_ws_created",
+    }
+)
+
+
+def _put_with_priority(
+    q: queue.Queue[dict[str, Any]],
+    data: dict[str, Any],
+    *,
+    critical: bool,
+) -> None:
+    """Put ``data`` onto ``q`` with selective drop-on-full semantics.
+
+    Critical events evict one oldest item from ``q`` if it's full,
+    then retry the put. Best-effort events drop themselves on full
+    (matches the historical ``contextlib.suppress(queue.Full)``
+    behaviour). The eviction is bounded — one drop is enough since
+    the queue was already at capacity before we tried.
+    """
+    if critical:
+        try:
+            q.put_nowait(data)
+        except queue.Full:
+            with contextlib.suppress(queue.Empty):
+                q.get_nowait()
+            with contextlib.suppress(queue.Full):
+                q.put_nowait(data)
+    else:
+        with contextlib.suppress(queue.Full):
+            q.put_nowait(data)
+
+
 # Cap on the per-turn assistant content accumulator. The accumulator
 # is piggybacked onto the ``ws_state:idle`` broadcast payload so the
 # cluster collector / dashboard can render the freshly-emitted assistant
@@ -246,14 +296,22 @@ class SessionUIBase:
         browser can validate it belongs to the pane's current
         workstream. Shallow-copies on stamp to avoid mutating a
         caller-owned dict.
+
+        Critical event types (see :data:`_CRITICAL_EVENT_TYPES`)
+        evict one oldest item from a full queue rather than dropping
+        themselves — a chatty mid-generation burst can fill the
+        500-slot per-tab queue, and silently dropping an
+        ``approve_request`` or ``intent_verdict`` because we couldn't
+        squeeze past a backlog of token-stream events would be a real
+        UX regression. Best-effort events drop on full as before.
         """
         if "ws_id" not in data:
             data = {**data, "ws_id": self.ws_id}
+        critical = data.get("type") in _CRITICAL_EVENT_TYPES
         with self._listeners_lock:
             snapshot = list(self._listeners)
         for lq in snapshot:
-            with contextlib.suppress(queue.Full):
-                lq.put_nowait(data)
+            _put_with_priority(lq, data, critical=critical)
 
     def _register_listener(
         self, maxsize: int = _DEFAULT_LISTENER_QUEUE_MAX
@@ -327,6 +385,11 @@ class SessionUIBase:
                 "always": bool(always),
             }
         )
+        # Kind-specific cross-stream broadcast — ConsoleCoordinatorUI
+        # overrides to push onto the cluster bus so a coord parent's
+        # tree UI clears the pending-approval pill in lockstep with
+        # the actual decision. Stage 3 Step 4.
+        self._broadcast_approval_resolved(approved, feedback, always=always)
         self._approval_event.set()
 
     @staticmethod
@@ -579,6 +642,20 @@ class SessionUIBase:
             "judge_pending": judge_pending,
         }
         self._enqueue(self._pending_approval)
+        # Cross-stream broadcast — push the items via the cluster bus
+        # so a coord parent's tree UI can render the inline approve/deny
+        # block without waiting for a bulk fetch. Without this, the
+        # bulk fetch races with this assignment: the state transition
+        # to ATTENTION fires upstream BEFORE approve_tools runs (see
+        # session.py:_emit_state("attention") preceding ui.approve_tools),
+        # so a bulk fetch landing in the ~50-200ms window between
+        # _emit_state and this point sees ``_pending_approval=None``
+        # and returns ``pending_approval_detail: null``. The 5s TTL
+        # then locks the coord row on a "loading" placeholder until
+        # the next state event triggers a refresh — which never comes
+        # while parked on _approval_event.wait. The push path
+        # eliminates the race.
+        self._broadcast_approve_request(self._pending_approval)
         if not self._approval_event.wait(timeout=self._APPROVAL_WAIT_TIMEOUT):
             # Approval timed out (e.g., user disconnected). Deny via
             # resolve_approval so verdicts and state are updated consistently.
@@ -632,6 +709,12 @@ class SessionUIBase:
                     del self._llm_verdicts[oldest_key]
                 self._llm_verdicts[call_id] = verdict
         self._enqueue({"type": "intent_verdict", **verdict})
+        # Kind-specific cross-stream broadcast — ConsoleCoordinatorUI
+        # overrides to push onto the cluster bus so a coord parent's
+        # tree UI sees the verdict without polling. Default is no-op
+        # (the per-ws ``_enqueue`` above already covers WebUI's own
+        # SSE listeners). Stage 3 Step 4.
+        self._broadcast_intent_verdict(verdict)
         self._persist_intent_verdict(verdict)
         # Decision check + either queue or flag-for-persist happen
         # under ONE lock acquisition so resolve_approval can't swap-
@@ -1351,6 +1434,42 @@ class SessionUIBase:
         Default: no-op. Subclasses override.
         """
 
+    def _broadcast_intent_verdict(self, verdict: dict[str, Any]) -> None:  # noqa: ARG002 — hook stub
+        """Fan an LLM intent-judge verdict out to the kind's transport.
+
+        Default: no-op. ``ConsoleCoordinatorUI`` overrides to push a
+        ``intent_verdict`` event onto the cluster bus so the parent
+        coordinator's tree UI can render the risk pill + verdict
+        result without polling. Stage 3 Step 4: hook only — the
+        cluster-bus event class lands in Step 5.
+        """
+
+    def _broadcast_approval_resolved(
+        self,
+        approved: bool,  # noqa: ARG002 — hook stub
+        feedback: str | None = None,  # noqa: ARG002 — hook stub
+        *,
+        always: bool = False,  # noqa: ARG002 — hook stub
+    ) -> None:
+        """Fan an ``approval_resolved`` decision out to the kind's transport.
+
+        Default: no-op. ``ConsoleCoordinatorUI`` overrides to push to
+        the cluster bus so the parent coordinator's tree UI can clear
+        the pending-approval pill in sync with the actual decision.
+        """
+
+    def _broadcast_approve_request(self, detail: dict[str, Any]) -> None:  # noqa: ARG002 — hook stub
+        """Fan an ``approve_request`` payload out to the kind's transport.
+
+        Default: no-op. ``WebUI`` and ``ConsoleCoordinatorUI`` override
+        to push the items list (the same dict that landed in
+        ``_pending_approval``) onto their respective transports. The
+        push path eliminates the bulk-fetch race that otherwise
+        leaves coord rows stuck on a loading placeholder when the
+        bulk fetch lands in the gap between the state transition to
+        ATTENTION and ``_pending_approval`` being set.
+        """
+
     # ------------------------------------------------------------------
     # State-broadcast snapshot helper
     # ------------------------------------------------------------------

turnstone/server.py

@@ -81,6 +81,7 @@ from turnstone.core.session_routes import (
 from turnstone.core.session_ui_base import (
     AutoApproveReason,
     SessionUIBase,
+    _put_with_priority,
     fire_judge_verdict_metric,
 )
 from turnstone.core.tools import TOOLS  # noqa: F401 — available for introspection
@@ -194,18 +195,14 @@ class WebUI(SessionUIBase):
             }
             if state == "idle":
                 event["content"] = payload["content"]
-            # Coord tree-UI renders inline approve/deny buttons off
-            # ``pending_approval_detail``; carrying it on the
-            # state-change broadcast lets the cluster bus update those
-            # buttons in lockstep with ``activity_state`` instead of
-            # forcing the browser to chase a separate dashboard fetch.
-            # Gated on existence so we don't pay the serializer's
-            # per-broadcast verdict-cache deepcopy on the common
-            # no-approval-pending path.
-            if self._pending_approval is not None:
-                detail = self.serialize_pending_approval_detail()
-                if detail is not None:
-                    event["pending_approval_detail"] = detail
+            # ``pending_approval_detail`` is NO LONGER piggybacked on
+            # state-change events (Stage 3 cleanup). Symmetric event
+            # flow now: initial approval items arrive via bulk fetch
+            # triggered by the ``activity_state="approval"`` transition,
+            # individual verdicts via the explicit
+            # ``intent_verdict`` event class, and resolution via
+            # ``approval_resolved``. Reducer no longer has to dedupe
+            # the piggyback path against the explicit one.
             try:
                 WebUI._global_queue.put_nowait(event)
             except queue.Full:
@@ -230,6 +227,82 @@ class WebUI(SessionUIBase):
                     }
                 )
 
+    def _broadcast_intent_verdict(self, verdict: dict[str, Any]) -> None:
+        """Send an LLM intent-judge verdict to the global SSE channel.
+
+        Stage 3 Step 5 — the cluster collector's ``_apply_delta``
+        forwards this verbatim to the cluster bus, where coord
+        adapters dispatch it as ``child_ws_intent_verdict`` for the
+        owning parent's tree UI. Unlike the existing
+        ``pending_approval_detail`` piggyback on ``ws_state``, this
+        fires WHENEVER a verdict lands — including the common case
+        where the judge daemon writes during ``attention`` with no
+        state transition to ride along on.
+        """
+        if WebUI._global_queue is not None:
+            # Critical event — evict an oldest item if the global
+            # queue is full rather than dropping the verdict.
+            _put_with_priority(
+                WebUI._global_queue,
+                {
+                    "type": "intent_verdict",
+                    "ws_id": self.ws_id,
+                    "verdict": verdict,
+                },
+                critical=True,
+            )
+
+    def _broadcast_approval_resolved(
+        self,
+        approved: bool,
+        feedback: str | None = None,
+        *,
+        always: bool = False,
+    ) -> None:
+        """Send an ``approval_resolved`` decision to the global SSE channel.
+
+        Clears the parent's pending-approval pill in lockstep with
+        the actual decision rather than waiting for the next
+        state-change piggyback.
+        """
+        if WebUI._global_queue is not None:
+            # Critical event — evict an oldest item if the global
+            # queue is full rather than dropping the resolution.
+            _put_with_priority(
+                WebUI._global_queue,
+                {
+                    "type": "approval_resolved",
+                    "ws_id": self.ws_id,
+                    "approved": approved,
+                    "feedback": feedback or "",
+                    "always": bool(always),
+                },
+                critical=True,
+            )
+
+    def _broadcast_approve_request(self, detail: dict[str, Any]) -> None:
+        """Send an ``approve_request`` payload to the global SSE channel.
+
+        Push path for the initial approval items so a coord parent's
+        tree UI can render the inline approve/deny block immediately
+        without waiting for a bulk-fetch round-trip. The bulk fetch
+        races with ``_pending_approval`` being set inside
+        ``approve_tools`` (the state transition to ATTENTION fires
+        upstream first); the push path eliminates that race entirely.
+        """
+        if WebUI._global_queue is not None:
+            # Critical event — evict an oldest item if the global
+            # queue is full rather than dropping the items.
+            _put_with_priority(
+                WebUI._global_queue,
+                {
+                    "type": "approve_request",
+                    "ws_id": self.ws_id,
+                    "detail": detail,
+                },
+                critical=True,
+            )
+
     # --- SessionUI protocol ---
     #
     # ``on_thinking_start`` / ``on_thinking_stop`` / ``on_reasoning_token``