Record the verified pose-capture recipe and the harness defect

MEMORY.md

@@ -11,12 +11,14 @@ Session memory for durable decisions, state, and gotchas. Keep concise, remove o
   direct `set_bone_pose_rotation` on HipL/KneeL/Spine/ShoulderL/ElbowL/Head changes rendered
   pixels; a bone *position* offset moves the head; a standalone probe running the same
   `_animate()` sequence renders the crouch. A staged in-game `film` frame also showed the arms
-  Standalone SceneTree probes (avatar under `root`, under a `Node3D` stage, and created
-  mid-`_process`) all render the seated crouch, so the rig and animation math are fine and the
-  harness capture path is the outlier. A headless test now locks the proxy→bone contract
-  (`test_pose_reaches_the_skeleton_bones`, suite 273). Next leads: reproduce the stage inside
-  `capture_npc_review.gd` (known-good captures) or capture via `SubViewport` +
-  `RenderingServer.force_draw()`.
+  Standalone SceneTree probes (avatar under `root`, under a `Node3D` stage, created
+  mid-`_process`) all render the seated crouch, and a `SubViewport` + `RenderingServer.force_draw()`
+  probe does too (window probes diff ~2,200-2,360; offscreen 602 at +2/+6/+12 frames, including a
+  re-created avatar). So the rig, the animation math and the offscreen path are sound; only
+  `capture_npc_pose.gd` still saves rest frames, which makes its phase/camera/force-draw ordering
+  the defect. A headless test locks the proxy→bone contract
+  (`test_pose_reaches_the_skeleton_bones`, suite 273). Rebuild the harness from the verified
+  recipe in `docs/3d-model-review.md` instead of iterating on the current phase machine.
 - **3D review baseline (2026-11-03):** `docs/3d-model-review.md` records the first skill-guided
   measurements: 46/46 assets clean, library 59,012 tris, characters 4,482–5,474 tris (68–69% of
   the 8,000 caps), `npc_base` only 1,204/2,488, 24-NPC 6.94 ms, and authored texel density