diff --git a/docs/superpowers/plans/2026-05-08-phase-n4-rendering-foundation.md b/docs/superpowers/plans/2026-05-08-phase-n4-rendering-foundation.md
index ba2337d..f31e820 100644
--- a/docs/superpowers/plans/2026-05-08-phase-n4-rendering-foundation.md
+++ b/docs/superpowers/plans/2026-05-08-phase-n4-rendering-foundation.md
@@ -837,6 +837,47 @@ writing a real bridge that shares index caches with our `DatCollection`.
 
 **No work for this task — skip and proceed to Task 7.**
 
+---
+
+### Adjustment 2 (2026-05-08): Task 9 routing reverted — tier decision belongs at spawn-callback layer
+
+**Discovered during Week 1 visual smoke test**: with flag on, characters /
+NPCs disappeared along with static scenery. Root cause: Task 9 routed
+**all** `InstancedMeshRenderer.EnsureUploaded` calls through
+`WbMeshAdapter.IncrementRefCount` and marked their cache entries with
+`WbManagedSentinel`. But `InstancedMeshRenderer` is used for both tiers
+in production:
+
+- **Atlas-tier** call sites: `_pendingCellMeshes` drain
+  ([GameWindow.cs:5137](../../../src/AcDream.App/Rendering/GameWindow.cs:5137)),
+  per-MeshRef GfxObj loop on `lb.Entities`
+  ([:5155](../../../src/AcDream.App/Rendering/GameWindow.cs:5155)).
+- **Per-instance-tier** call sites: per-part loop in spawn handling
+  ([:2302](../../../src/AcDream.App/Rendering/GameWindow.cs:2302)) — this is
+  character / creature rendering driven by server `CreateObject`.
+
+The renderer is **tier-blind by design**: it doesn't know spawn source.
+Putting routing logic there violates separation of concerns. The spec's
+Data-Flow section already specifies the right placement — routing happens
+at the **spawn-callback layer**:
+
+- `LandblockSpawnAdapter.OnLandblockLoaded(...)` (Task 11) calls
+  `IncrementRefCount` per unique GfxObj — atlas-tier only.
+- `EntitySpawnAdapter.OnCreate(entity)` (Task 17) routes through
+  per-instance path (`TextureCache.GetOrUploadWithPaletteOverride`) —
+  never calls `IncrementRefCount` for atlas.
+
+**Resolution:** reverted Task 9's renderer-level routing. Removed the
+sentinel logic and the 4 sentinel-skip checks in
+`InstancedMeshRenderer`. **Kept** the `_wbMeshAdapter` constructor
+parameter (unused for now) so `GameWindow.cs` doesn't shift when
+later tasks need adapter access. Kept all the real WB pipeline
+construction in `WbMeshAdapter` (verified working under flag-off).
+
+**Week 1 endpoint shifts:** "WB infrastructure constructed; flag-on and
+flag-off visually identical." Routing arrives in Week 2 (Task 11) at
+the correct layer. Smoke verification is now: flag-on === flag-off.
+
 ### Task 6 (original — kept for history)
 
 **Files:**
diff --git a/src/AcDream.App/Rendering/InstancedMeshRenderer.cs b/src/AcDream.App/Rendering/InstancedMeshRenderer.cs
index 2ba5093..5b0c9eb 100644
--- a/src/AcDream.App/Rendering/InstancedMeshRenderer.cs
+++ b/src/AcDream.App/Rendering/InstancedMeshRenderer.cs
@@ -35,19 +35,16 @@ public sealed unsafe class InstancedMeshRenderer : IDisposable
     private readonly TextureCache _textures;
 
     /// <summary>
-    /// Optional WB adapter. When non-null and <see cref="WbFoundationFlag.IsEnabled"/>,
-    /// <see cref="EnsureUploaded"/> hands the GfxObj ref to the WB pipeline instead of
-    /// uploading into our own VAO pool. The draw loop skips sentinel entries — Task 22's
-    /// WbDrawDispatcher will eventually draw them.
+    /// Optional WB adapter. Held but currently unused — Phase N.4 Adjustment 2
+    /// (2026-05-08) reverted Task 9's renderer-level routing. Tier-routing decisions
+    /// (atlas vs per-instance) belong at the spawn-callback layer (Task 11
+    /// LandblockSpawnAdapter for atlas-tier; Task 17 EntitySpawnAdapter for
+    /// per-instance), not in the renderer which is intentionally tier-blind. The
+    /// constructor parameter is preserved so GameWindow's wire-up doesn't shift
+    /// when later tasks need adapter access.
     /// </summary>
     private readonly WbMeshAdapter? _wbMeshAdapter;
 
-    // Sentinel: a GfxObj that has been handed to the WB pipeline gets this list
-    // stored in _gpuByGfxObj. The Draw loop recognises it by reference identity
-    // (object.ReferenceEquals) and skips it — no legacy VAO draw for WB-managed
-    // objects until Task 22 wires up WbDrawDispatcher.
-    private static readonly List<SubMeshGpu> WbManagedSentinel = new(0);
-
     // One GPU bundle per unique GfxObj id. Each GfxObj can have multiple sub-meshes.
     private readonly Dictionary<uint, List<SubMeshGpu>> _gpuByGfxObj = new();
 
@@ -100,17 +97,11 @@ public sealed unsafe class InstancedMeshRenderer : IDisposable
         if (_gpuByGfxObj.ContainsKey(gfxObjId))
             return;
 
-        // Phase N.4 Task 9: when the WB foundation flag is on and we have an
-        // adapter, hand this GfxObj to the WB pipeline instead of uploading our
-        // own VAO. The sentinel entry marks "this GfxObj lives in WB now" so the
-        // draw loop knows to skip it. Task 22's WbDrawDispatcher will draw them.
-        if (WbFoundationFlag.IsEnabled && _wbMeshAdapter is not null)
-        {
-            _wbMeshAdapter.IncrementRefCount(gfxObjId);
-            _gpuByGfxObj[gfxObjId] = WbManagedSentinel;
-            return;
-        }
-
+        // Phase N.4 Adjustment 2 (2026-05-08): renderer is tier-blind. Tier-routing
+        // (atlas vs per-instance) lives at the spawn-callback layer (Tasks 11 + 17),
+        // not here. Smoke-test of the original Task 9 routing showed it caught
+        // characters / NPCs (server-spawned, per-instance tier) along with static
+        // scenery, because EnsureUploaded is called from both spawn paths.
         var list = new List<SubMeshGpu>(subMeshes.Count);
         foreach (var sm in subMeshes)
             list.Add(UploadSubMesh(sm));
@@ -245,11 +236,6 @@ public sealed unsafe class InstancedMeshRenderer : IDisposable
             if (!_gpuByGfxObj.TryGetValue(key.GfxObjId, out var subMeshes))
                 continue;
 
-            // WB-managed GfxObjs have a sentinel entry; Task 22 (WbDrawDispatcher)
-            // will draw them. Skip here to avoid drawing with stale/null VAO data.
-            if (object.ReferenceEquals(subMeshes, WbManagedSentinel))
-                continue;
-
             bool hasOpaqueSubMesh = false;
             foreach (var sub in subMeshes)
             {
@@ -325,10 +311,6 @@ public sealed unsafe class InstancedMeshRenderer : IDisposable
             if (!_gpuByGfxObj.TryGetValue(key.GfxObjId, out var subMeshes))
                 continue;
 
-            // WB-managed GfxObjs — skip; Task 22 will draw them.
-            if (object.ReferenceEquals(subMeshes, WbManagedSentinel))
-                continue;
-
             bool hasTranslucentSubMesh = false;
             foreach (var sub in subMeshes)
             {
@@ -458,9 +440,6 @@ public sealed unsafe class InstancedMeshRenderer : IDisposable
                 {
                     if (!_gpuByGfxObj.TryGetValue(meshRef.GfxObjId, out var cachedMeshes))
                         continue;
-                    // WB-managed GfxObjs don't go through our instance pipeline.
-                    if (object.ReferenceEquals(cachedMeshes, WbManagedSentinel))
-                        continue;
 
                     var model = meshRef.PartTransform * entityRoot;
 
@@ -565,11 +544,6 @@ public sealed unsafe class InstancedMeshRenderer : IDisposable
     {
         foreach (var subs in _gpuByGfxObj.Values)
         {
-            // WB-managed entries use the sentinel — no GL resources to free here;
-            // ObjectMeshManager owns those resources.
-            if (object.ReferenceEquals(subs, WbManagedSentinel))
-                continue;
-
             foreach (var sub in subs)
             {
                 _gl.DeleteBuffer(sub.Vbo);