diff --git a/src/AcDream.Core/Physics/PhysicsEngine.cs b/src/AcDream.Core/Physics/PhysicsEngine.cs
index d23a000..f225b4d 100644
--- a/src/AcDream.Core/Physics/PhysicsEngine.cs
+++ b/src/AcDream.Core/Physics/PhysicsEngine.cs
@@ -597,33 +597,36 @@ public sealed class PhysicsEngine
         if (isOnGround)
             transition.ObjectInfo.State |= ObjectInfoState.Contact | ObjectInfoState.OnWalkable;
 
-        // K-fix7 (2026-04-26): only seed the contact plane when the body
-        // is actually grounded. Pre-seeding while AIRBORNE caused
-        // AdjustOffset's "Have a contact plane / Moving away from plane"
-        // branch to fire on every jump step — which calls
-        // Plane::snap_to_plane on the offset and ZEROES the Z component,
-        // killing all upward jump motion (the body's Z velocity stayed
-        // ~9 m/s but Position.Z never advanced because every step's
-        // offset got snapped flat). Retail's CTransition::init at
-        // 0x509dd0 (named-retail line 271954) explicitly clears
-        // contact_plane_valid = 0 at the start of every transition
-        // resolve, then ValidateWalkable re-establishes it during the
-        // sweep when the sphere bottom is within EPSILON of the terrain
-        // plane — so for grounded motion the plane is set fresh every
-        // resolve, and for airborne motion no plane interferes.
+        // A6.P3 slice 2 (2026-05-22) — DO NOT seed ContactPlane at
+        // transition start. Retail's CTransition::init at 0x509dd0
+        // (acclient_2013_pseudo_c.txt:271954) explicitly clears
+        // contact_plane_valid = 0 — retail does NOT seed. Cross-tick
+        // CP retention flows via:
+        //   - Mechanism A: BSPQuery.FindCollisions Path-6 land write
+        //     (writes ContactPlane during the sub-step loop when the
+        //     sphere ground-contacts a floor poly)
+        //   - Mechanism B: Transition.ValidateTransition LKCP restore
+        //     (slice 1 commit 5aba071; restores ContactPlane from
+        //     LastKnownContactPlane post-sub-step when sphere close to
+        //     LKCP)
+        //   - Body persist below: at transition end, body.ContactPlane
+        //     gets set from ci.ContactPlane OR ci.LastKnownContactPlane
+        //     so the NEXT tick's body state is preserved.
         //
-        // We KEEP the seeding when isOnGround for slope-walking
-        // continuity (the original concern that motivated the seed) —
-        // walking up a hill needs the previous step's slope to project
-        // movement properly. Airborne / jumping must start with no
-        // plane so AdjustOffset preserves Z.
-        if (isOnGround && body is not null && body.ContactPlaneValid)
-        {
-            transition.CollisionInfo.SetContactPlane(
-                body.ContactPlane,
-                body.ContactPlaneCellId,
-                body.ContactPlaneIsWater);
-        }
+        // Cost: AdjustOffset on sub-step 1 of each tick has no CP and
+        // takes the "no contact plane" path (TransitionTypes.cs:2329-2334)
+        // — slide-projects only, no slope-snap. Sub-steps 2+ pick up CP
+        // from Mechanism A / B as normal. Imperceptible in practice.
+        //
+        // Removing this seed closes issue #96 (per-tick CP-write blowup;
+        // 99.3% of post-slice-1 CP writes came from the seed) and is
+        // hypothesized to close issue #97 (phantom collisions +
+        // fall-through; suspected stale-CP slope-snap was the cause).
+        //
+        // Previous K-fix7 seed (2026-04-26 to slice-1 ship) deliberately
+        // diverged from retail for "slope-walking continuity" but the
+        // tradeoff (per-tick CP writes, stale-CP slope-snap) was not
+        // worth it.
 
         // Retail CPhysicsObj::get_object_info also seeds SlidingNormal when
         // transient_state has bit 2 set. This matters for one-step/frame hits: