diff --git a/CLAUDE.md b/CLAUDE.md
index 328264c..75ff1fe 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -164,10 +164,22 @@ The triangle-boundary Z bug cost 5 failed fix attempts from guessing.
 The animation frame-swap bug cost 4 failed attempts. Every time we
 checked the decompiled code first, we got it right on the first try.
 **Now we have named retail symbols too — Step 0 cuts most lookups
-from 30 minutes to 5 seconds.**
+from 30 minutes to 5 seconds. And as of 2026-04-30, when "what does
+retail actually DO at runtime?" is the question and decomp alone
+isn't enough, attach cdb to a live retail client (Step -1).**
 
 ### For each new feature or bug fix:
 
+-1. **ATTACH cdb TO RETAIL (when behavior is the question, not code).**
+   For "what does retail actually DO frame-by-frame?" questions —
+   wedges, weird animation flicker, geometry-specific bugs, anything
+   where the decomp is correct but it's not clear how it produces the
+   visible behavior — **don't guess; attach the Windows debugger to
+   a live retail client and trace it.** See "Retail debugger toolchain"
+   below for setup. We discovered the steep-roof wedge had a 30Hz
+   physics-tick cause this way; would have taken weeks of guessing
+   without the trace.
+
 0. **GREP NAMED FIRST.** Before any decompilation work, search
    `docs/research/named-retail/acclient_2013_pseudo_c.txt` by
    `class::method` name. 99.6% of functions have real names from the
@@ -249,6 +261,124 @@ Before marking any phase as done:
 - [ ] Roadmap updated
 - [ ] Memory updated if there's a durable lesson
 
+## Retail debugger toolchain (live runtime trace)
+
+**When the question is "what does retail actually DO frame-by-frame?"**
+the decomp alone is often not enough — code paths interact with state
+(LastKnownContactPlane, transient flags, accumulated counters) in ways
+that aren't obvious from reading. As of 2026-04-30 we have a working
+toolchain to attach Windows' console debugger (cdb.exe) to a live
+retail acclient.exe with full PDB symbols and capture state at any
+breakpoint. **Use this when guessing has failed twice in a row.**
+
+### What we have
+
+- **Matching binary**: `C:\Turbine\Asheron's Call\acclient.exe`
+  v11.4186 (linker timestamp `2013-09-06 00:17:42 UTC`,
+  CodeView GUID `9e847e2f-777c-4bd9-886c-22256bb87f32`). Pairs
+  exactly with our `refs/acclient.pdb`.
+- **Debugger**: `cdb.exe` (console WinDbg) at
+  `C:\Program Files (x86)\Windows Kits\10\Debuggers\x86\cdb.exe`.
+  Install via Microsoft Store WinDbg (~50 MB). 32-bit version is
+  required for acclient.exe.
+- **PDB**: `refs/acclient.pdb` (29 MB, Sept 2013 EoR build).
+  18,366 named functions + 5,371 named struct types resolve.
+- **Symbol verifier**: `tools/pdb-extract/check_exe_pdb.py <exe>`
+  reads any acclient.exe and prints whether it pairs with our PDB
+  (`MATCH` / `MISMATCH (expected GUID = ...)`). Always run this on
+  a candidate binary BEFORE attaching.
+- **PDB metadata dumper**: `tools/pdb-extract/dump_pdb_info.py refs/acclient.pdb`
+  prints the PDB's expected timestamp + GUID + age. Use to figure
+  out which build to look for if the chain ever breaks.
+
+### Workflow
+
+1. **Verify the binary matches the PDB:**
+   ```bash
+   py tools/pdb-extract/check_exe_pdb.py "C:/Turbine/Asheron's Call/acclient.exe"
+   ```
+   Expect: `=== MATCH: this exe pairs with our acclient.pdb ===`
+
+2. **Have the user launch retail client** and connect to local ACE.
+   Retail must already be in-world before attaching.
+
+3. **Write a `.cdb` script** that arms breakpoints with non-blocking
+   actions (count + log + `gc`). Pattern:
+   ```
+   .logopen <output-path>
+   .sympath C:\Users\erikn\source\repos\acdream\refs
+   .symopt+ 0x40
+   .reload /f acclient.exe
+
+   r $t0 = 0
+   bp acclient!CTransition::transitional_insert "r $t0 = @$t0 + 1; .if (@$t0 % 5000 == 0) { .printf \"...\" }; .if (@$t0 >= 30000) { qd } .else { gc }"
+   bp acclient!OBJECTINFO::kill_velocity "r $t1 = @$t1 + 1; gc"
+   ...
+   g
+   ```
+   `gc` = "go conditional" (continue without breaking). Auto-detach
+   via `qd` after a hit-count threshold to avoid manual cleanup.
+
+4. **Launch cdb in the background** via a PowerShell wrapper:
+   ```powershell
+   & "C:\Program Files (x86)\Windows Kits\10\Debuggers\x86\cdb.exe" `
+       -pn acclient.exe -cf <script>.cdb *>&1 |
+       Tee-Object -FilePath <log>
+   ```
+
+5. **User reproduces the scenario** in the retail client window
+   (jump on roof, hit wall, etc.). Breakpoints fire, log fills.
+
+6. **cdb auto-detaches** when the threshold breakpoint fires `qd`.
+   Retail keeps running unaffected. Read the log offline.
+
+### Known watchouts
+
+- **PDB function names use snake_case for some classes**
+  (BSPTREE, CTransition, OBJECTINFO, COLLISIONINFO, SPHEREPATH) and
+  **PascalCase for others** (CPhysicsObj). The Binary Ninja decomp
+  shows snake_case for everything; the PDB has Turbine's actual
+  PascalCase for CPhysicsObj. Always look up symbols with `x` first
+  to find the actual name.
+
+- **`bp acclient!Class::method`** sets a breakpoint by symbol. The
+  cdb command parser splits on `;`, so don't put `;` inside the
+  action string — use newlines or escape carefully.
+
+- **Symbol path: do NOT use `.sympath srv*<server>;<local>`** — the
+  `;` is a cdb command separator, gets split. Use `.sympath <local>`
+  (no symbol server, just our refs/) since we don't need Microsoft
+  system DLL symbols.
+
+- **Killing cdb kills the debuggee.** Use `qd` (quit detached) inside
+  a breakpoint action to detach cleanly. `Stop-Process cdb` will
+  take the retail client down with it.
+
+- **High breakpoint hit rates produce game lag.** Each breakpoint hit
+  traps the process briefly. For frequent functions
+  (transitional_insert at ~10K/sec) the cumulative cost is enough to
+  make retail feel sluggish. Mitigate by setting a tight auto-detach
+  threshold (e.g., 30,000 hits) and/or moving counters to less-frequent
+  functions.
+
+- **acclient.exe is 32-bit + uses thiscall.** When dumping struct
+  fields in breakpoint actions, `this` is in `ecx`. Use cdb's
+  `dt acclient!ClassName @ecx` for full struct dump.
+
+### When NOT to use this
+
+- **Pure code-port questions** — the decomp at `docs/research/named-retail/`
+  has the answer. Don't waste time on cdb if `grep` is enough.
+- **Visual / rendering bugs** — debugger doesn't help with shaders or
+  framebuffers; use RenderDoc or similar.
+- **Network protocol questions** — `holtburger` references + ACE source
+  + Wireshark are the right tools, not cdb.
+
+This toolchain was used to settle the L.5 steep-roof investigation:
+30Hz physics tick (vs our 60Hz), `kill_velocity` gating,
+`set_collide` rate per minute. See commit history around 2026-04-30
+for the trace data and the decisions it drove.
+
 ## Subagent policy
 
 Subagents are the primary tool for saving parent-context and keeping one
diff --git a/docs/ISSUES.md b/docs/ISSUES.md
index 6d12894..2b625f0 100644
--- a/docs/ISSUES.md
+++ b/docs/ISSUES.md
@@ -196,11 +196,44 @@ before `SPHEREPATH::precipice_slide`; edge-slide `Slid` / `Adjusted` results
 now feed the `TransitionalInsert` retry loop instead of being reverted by outer
 validation, and a synthetic diagonal terrain-boundary test covers tangent
 motion. `ACDREAM_DUMP_EDGE_SLIDE=1` now reports whether a failed step-down had
-polygon context. Remaining gaps: live visual confirmation of the retry-loop
-fix, real-DAT building-edge fixtures, fuller `cliff_slide` coverage, and
-`NegPolyHit` dispatch. Named retail anchors include `CTransition::edge_slide`,
-`CTransition::cliff_slide`, `SPHEREPATH::precipice_slide`, and
-`SPHEREPATH::step_up_slide`.
+polygon context.
+
+**L.4/L.5 update 2026-04-30:** A retail debugger trace (cdb attached to
+v11.4186 acclient.exe — see #35) confirmed that retail does NOT wedge
+on the steep-roof scenario that produces the wedge in our acdream port.
+Three concrete findings:
+1. Retail's `OBJECTINFO::kill_velocity` rarely fires in normal play —
+   gated on `last_known_contact_plane_valid`, which our L.2.4 proximity
+   guard tends to clear before steep-poly hits land. Retail trace: 0
+   kill_velocity hits across 40,960 update_object calls. Our Phase 3
+   reset path now matches retail's gate (only kills when valid).
+2. Retail integrates physics at 30Hz (`MinQuantum = 1/30 s`); render is
+   60+ Hz. UpdatePhysicsInternal/update_object ratio = 0.61. We
+   ported this gate as L.5 in `PlayerMovementController` via
+   `_physicsAccum`. Render still runs at 60+ Hz; only the physics
+   integration step is 30Hz.
+3. The remaining wedge cause — body's pre-position drifts to the
+   polygon's tangent and gravity's tangent component into surface
+   produces a stable retain-collide-revert loop — is a downstream
+   consequence of retail's grounded-on-steep escape chain
+   (`step_sphere_up` → `step_up_slide` → `cliff_slide`) being
+   incompletely ported. Live test confirmed retail-strict Path 6
+   produces "lands on roof in falling animation, can't slide off"
+   half-state because that chain doesn't produce smooth descent.
+
+**Pragmatic ship-state:** BSPQuery Path 6 keeps the L.4 slide-tangent
+deviation (project-along-steep-face-and-return-Slid) for steep-poly
+airborne hits. It produces user-acceptable "slide off the roof"
+behavior at the cost of departing from retail's Path 6 → SetCollide →
+Path 4 → Phase 3 reset chain. Retail-strict requires the
+step_up_slide / cliff_slide audit below; until that lands, slide-tangent
+is the right deviation.
+
+Remaining gaps: real-DAT building-edge fixtures, fuller `cliff_slide`
+coverage, `NegPolyHit` dispatch, and the retail-strict
+step_up_slide / cliff_slide audit (filed for follow-up). Named retail
+anchors include `CTransition::edge_slide`, `CTransition::cliff_slide`,
+`SPHEREPATH::precipice_slide`, and `SPHEREPATH::step_up_slide`.
 
 **Files:** `src/AcDream.Core/Physics/TransitionTypes.cs`,
 `src/AcDream.Core/Physics/BSPQuery.cs`,
@@ -214,6 +247,52 @@ cliff/precipice slide, failed step-up/step-down, and the jump-clears-edge case.
 
 ---
 
+## #35 — [DONE 2026-04-30] Retail debugger toolchain (cdb + PDB GUID matching)
+
+**Status:** DONE
+**Severity:** N/A (infrastructure)
+**Filed + closed:** 2026-04-30
+**Component:** tooling / research
+
+**Description:** When the question is "what does retail actually DO at
+runtime?" — wedges, animation flicker, geometry-specific bugs where the
+decomp is correct but the visible behavior is mysterious — there was no
+way to attach a debugger to a live retail acclient.exe and trace it.
+This issue tracks the toolchain that closed that gap.
+
+**What shipped:**
+- **`tools/pdb-extract/check_exe_pdb.py`** — reads any PE's CodeView entry
+  and reports `MATCH` / `MISMATCH (expected GUID = …)` against our
+  `refs/acclient.pdb`. Always run before attaching cdb.
+- **`tools/pdb-extract/dump_pdb_info.py`** — dumps a PDB's expected
+  build timestamp + GUID + age. Used to figure out which acclient.exe
+  build pairs with our PDB (answer: v11.4186, Sept 2013 EoR).
+- **CLAUDE.md "Retail debugger toolchain" section** — full workflow:
+  cdb path, sample `.cdb` script, PowerShell wrapper pattern, watchouts
+  (PDB name conventions, `;` parsing, kill-target-on-detach behavior,
+  high-hit-rate lag).
+- **Step `-1` added to the development workflow** — "ATTACH cdb TO
+  RETAIL (when behavior is the question, not code)". Tells future
+  sessions: when guessing has failed twice in a row, don't keep guessing.
+
+**Discoveries this toolchain enabled (closed in same session):**
+- Retail integrates physics at 30Hz (`UpdatePhysicsInternal/update_object`
+  ratio = 0.61). Drove the L.5 fix in PlayerMovementController.
+- `OBJECTINFO::kill_velocity` rarely fires in normal play (gated on
+  last_known_contact_plane_valid). Our acdream port now matches.
+- Retail does NOT wedge on the steep-roof scenario. Confirmed our L.4
+  slide-tangent deviation in Path 6 is necessary until the retail
+  step_up_slide / cliff_slide chain audit lands.
+
+**Files:** `tools/pdb-extract/check_exe_pdb.py`,
+`tools/pdb-extract/dump_pdb_info.py`, `CLAUDE.md`,
+`memory/project_retail_debugger.md`.
+
+**Acceptance:** Future sessions can attach cdb to a live retail client
+in under 5 minutes by following the CLAUDE.md workflow.
+
+---
+
 ## #33 — Live entity collision shape collapses to one cylinder
 
 **Status:** OPEN
diff --git a/src/AcDream.App/Input/PlayerMovementController.cs b/src/AcDream.App/Input/PlayerMovementController.cs
index 53c0225..e4fc8ec 100644
--- a/src/AcDream.App/Input/PlayerMovementController.cs
+++ b/src/AcDream.App/Input/PlayerMovementController.cs
@@ -172,6 +172,32 @@ public sealed class PlayerMovementController
     public const float HeartbeatInterval = 0.2f;  // 200ms
     public bool HeartbeatDue { get; private set; }
 
+    // L.5 retail physics-tick gate (2026-04-30).
+    //
+    // Retail's CPhysicsObj::update_object subdivides per-frame dt into
+    // MinQuantum (1/30s) sized integration steps, SKIPPING entirely when
+    // accumulated dt is below MinQuantum. The retail debugger trace
+    // confirmed this: UpdatePhysicsInternal fires only ~61% as often as
+    // update_object — i.e., retail's effective physics tick rate is 30Hz
+    // even when the renderer runs at 60+Hz.
+    //
+    // Without this gate our acdream integrates at the full render rate
+    // (60+Hz), which compresses bounce-energy / gravity-tangent
+    // accumulation into half the time. Per-frame V grows ~2x faster than
+    // retail's. On a steep-slope tangent that produces the wedge: V grows
+    // tangent + huge while position reverts each frame, body locks in
+    // place. Retail's slower integration cadence (and larger per-tick
+    // position deltas) lets the body geometrically escape the tangent.
+    //
+    // Source: retail debugger trace 2026-04-30
+    //   update_object       = 40,960 calls
+    //   UpdatePhysicsInternal = 25,087 calls   (61%)
+    //   ratio implies 39% of frames return early via the MinQuantum gate.
+    //
+    // ACE: PhysicsObj.UpdateObject (Physics.cs).
+    // Named-retail: CPhysicsObj::update_object (acclient_2013_pseudo_c.txt:283950).
+    private float _physicsAccum;
+
     public PlayerMovementController(PhysicsEngine physics)
     {
         _physics = physics;
@@ -411,9 +437,34 @@ public sealed class PlayerMovementController
         }
 
         // ── 4. Integrate physics (gravity, friction, sub-stepping) ────────────
+        //
+        // L.5 retail-physics-tick gate (2026-04-30): retail's CPhysicsObj::
+        // update_object skips integration when accumulated dt is below
+        // MinQuantum (1/30 s). Effective physics rate is 30 Hz even at 60+ Hz
+        // render. We accumulate per-frame dt and only integrate (with the
+        // accumulated dt) when the threshold is reached. See _physicsAccum
+        // declaration for the full retail trace evidence.
         var preIntegratePos = _body.Position;
-        _body.calc_acceleration();
-        _body.UpdatePhysicsInternal(dt);
+        _physicsAccum += dt;
+
+        if (_physicsAccum > PhysicsBody.HugeQuantum)
+        {
+            // Stale frame (debugger break, GC pause). Discard accumulated dt.
+            _physicsAccum = 0f;
+        }
+        else if (_physicsAccum >= PhysicsBody.MinQuantum)
+        {
+            // Integrate accumulated dt, clamped to MaxQuantum so a long
+            // pause doesn't produce one giant integration step.
+            float tickDt = MathF.Min(_physicsAccum, PhysicsBody.MaxQuantum);
+            _body.calc_acceleration();
+            _body.UpdatePhysicsInternal(tickDt);
+            _physicsAccum -= tickDt;
+        }
+        // Else: dt below MinQuantum threshold — skip integration. Position
+        // and velocity remain unchanged; Resolve below runs as a zero-distance
+        // sphere sweep (no collision possible) and the rest of the frame
+        // (motion commands, animation, return) runs normally.
         var postIntegratePos = _body.Position;
 
         // ── 5. Collision resolution via CTransition sphere-sweep ─────────────
@@ -520,32 +571,6 @@ public sealed class PlayerMovementController
                 ? !(prevOnWalkable && nowOnWalkable)
                 : (!prevOnWalkable && !nowOnWalkable);
 
-            // L.4-steep-landing-bounce (2026-04-30): also bounce on
-            // landing IF the contact surface is upward-facing but
-            // steeper than walkable (FloorZ ≈ 49°). Per retail and the
-            // user's expectation: jumping onto a steep roof should NOT
-            // result in a landing — the player should bounce off, keep
-            // the falling animation, and slide off.
-            //
-            // Without this: the L.3a base rule suppresses the bounce on
-            // landing transitions (prev air → now ground) to avoid
-            // micro-bounce on flat terrain, but that suppression also
-            // sticks the player to too-steep roofs they shouldn't land
-            // on. This carve-out re-enables the bounce specifically for
-            // steep upward-facing surfaces.
-            //
-            // Range `0 < N.Z < FloorZ` means "facing upward but too
-            // steep" — excludes walls (N.Z ≈ 0) which are handled by the
-            // existing prevAirborne+nowAirborne rule, and ceilings
-            // (N.Z < 0) which the body shouldn't bounce off the same way.
-            if (!applyBounce
-                && resolveResult.CollisionNormalValid
-                && resolveResult.CollisionNormal.Z > 0f
-                && resolveResult.CollisionNormal.Z < PhysicsGlobals.FloorZ)
-            {
-                applyBounce = true;
-            }
-
             // L.4-diag (2026-04-30): per-frame bounce trace for steep-roof bug.
             bool diagSteep = Environment.GetEnvironmentVariable("ACDREAM_DUMP_STEEP_ROOF") == "1";
             if (diagSteep && resolveResult.CollisionNormalValid)
diff --git a/src/AcDream.Core/Physics/BSPQuery.cs b/src/AcDream.Core/Physics/BSPQuery.cs
index cb49f9f..0cb17e4 100644
--- a/src/AcDream.Core/Physics/BSPQuery.cs
+++ b/src/AcDream.Core/Physics/BSPQuery.cs
@@ -1473,6 +1473,27 @@ public static class BSPQuery
             if (changed && hitPoly is not null)
             {
                 // ACE: var offset = LocalToGlobalVec(validPos.Center - localSphere.Center) * scale
+                //
+                // L.4 retail-strict (2026-04-30): the FloorZ gate previously
+                // here was REMOVED after the retail debugger trace + acdream
+                // wedge analysis showed it was preventing the natural escape
+                // path. Retail's flow:
+                //   Frame N: airborne sphere hits steep poly. Path 4 commits
+                //     ContactPlane on the steep poly (LandingZ ≈ 0.087 is
+                //     permissive enough — even 49°+ slopes pass).
+                //   Frame N+1: body now grounded with Contact + steep
+                //     ContactPlane. OnWalkable cleared by FloorZ test
+                //     downstream. Resolver fires Path 5 (Contact branch)
+                //     instead of Path 6. step_sphere_up tries to step over,
+                //     fails, falls back to step_up_slide → clears Contact,
+                //     slides sphere laterally along StepUpNormal.
+                //   Frame N+2: body airborne with lateral V from the slide.
+                //     Gravity takes over, body falls off the slope.
+                //
+                // Without this Path-4 commit, the body NEVER gets Contact
+                // set, Path 5 never fires, step_up_slide never runs, and the
+                // body wedges in airborne-collision-revert-loop with V at
+                // MaxVelocity tangent to the surface (live wedge 2026-04-30).
                 var localOffset = validPos.Center - sphere0.Center;
                 var worldOffset = L2W(localOffset) * scale;
                 path.AddOffsetToCheckPos(worldOffset);
@@ -1573,35 +1594,39 @@ public static class BSPQuery
 
                 var worldNormal0 = L2W(hitPoly0!.Plane.Normal);
 
-                // L.4-reject-steep-landing (2026-04-30): if the polygon is
-                // too steep to walk on (worldNormal.Z < FloorZ ≈ 0.6642),
-                // do NOT enter the SetCollide → Path-4 → SetContactPlane
-                // landing path. That path commits the player to the
-                // surface (sets ContactPlane), which sticks them to the
-                // steep roof in a falling animation.
+                // L.4 slide-tangent for steep airborne hits (2026-04-30).
                 //
-                // Instead, treat the steep-poly hit as a wall slide:
-                // project the move along the steep face (remove the
-                // into-wall component), set CollisionNormal +
+                // For polygons too steep to walk on (worldNormal.Z < FloorZ),
+                // skip the SetCollide → Path-4 → ContactPlane landing chain.
+                // That chain commits the body to the steep surface, leading
+                // to the "stuck in falling animation on the roof" bug — once
+                // grounded with a steep ContactPlane, our step_up_slide /
+                // cliff_slide / edge_slide chain can't produce smooth
+                // descent and the body wedges or "falls a bit at a time"
+                // when bumped.
+                //
+                // Instead: project the move along the steep face (remove
+                // the into-wall displacement), set CollisionNormal +
                 // SlidingNormal, return Slid. Same shape as Path 5's
                 // step-up fallback (line 1545-1547) and CylinderCollision
-                // (TransitionTypes.cs:1518-1522). The position is updated
-                // in-place; on the next resolver iteration the sphere is
+                // (TransitionTypes.cs:1518-1522). Position is updated in-
+                // place; on the next resolver iteration the sphere is
                 // outside the poly, FindCollisions returns OK, and
                 // ValidateTransition commits the new position. Body stays
-                // airborne, falling animation continues, gravity pulls
-                // down the slope.
+                // airborne, falling animation continues, and gravity's
+                // tangent component drifts the body downhill until it
+                // slides off the slope's edge.
                 //
-                // CRITICAL: this MUST happen before path.SetCollide(...)
-                // is called. Once Collide=true is set, TransitionalInsert
-                // Phase 3 either commits via ContactPlane+Placement (the
-                // walkable case, OK on shallow slopes) or RestoreCheckPos
-                // (the reset case, when ContactPlaneValid is false). The
-                // reset path REVERTS our slide and freezes the body.
-                //
-                // Per user 2026-04-30:
-                // "I jump up, I land on it. It should not even let me
-                //  land, should just slide with a falling animation."
+                // This is a deliberate deviation from retail (retail uses
+                // SetCollide unconditionally and lets find_walkable +
+                // step_up_slide produce the slide). Validated against
+                // retail debugger trace 2026-04-30: retail body did not
+                // wedge; our retail-faithful port DID wedge because we're
+                // missing implementation details of the step_up_slide /
+                // cliff_slide chain on grounded-steep movement. The
+                // slide-tangent here produces user-acceptable behavior
+                // (slides off naturally) while the deeper chain port is
+                // researched. Filed as L.5+ followup for retail-strict.
                 if (worldNormal0.Z < PhysicsGlobals.FloorZ)
                 {
                     Vector3 currWorld = path.GlobalCurrCenter[0].Origin;
@@ -1613,26 +1638,11 @@ public static class BSPQuery
 
                     collisions.SetCollisionNormal(worldNormal0);
                     collisions.SetSlidingNormal(worldNormal0);
-
-                    if (Environment.GetEnvironmentVariable("ACDREAM_DUMP_STEEP_ROOF") == "1")
-                    {
-                        Console.WriteLine(
-                            $"[steep-roof] PATH6-STEEP-SLIDE N=({worldNormal0.X:F2},{worldNormal0.Y:F2},{worldNormal0.Z:F2}) " +
-                            $"FloorZ={PhysicsGlobals.FloorZ:F2} " +
-                            $"diff={diff:F3} " +
-                            $"newCheckPos=({path.CheckPos.X:F2},{path.CheckPos.Y:F2},{path.CheckPos.Z:F2})");
-                    }
                     return TransitionState.Slid;
                 }
 
-                // ─── SetCollide response ─────────────────────────────────
-                // Airborne sphere hits a shallow polygon (walkable surface).
-                // Call SetCollide so TransitionalInsert's Collide branch
-                // re-tests as Placement to confirm we can land on it.
-                //
-                // ACE: BSPTree.find_collisions default branch → SpherePath.SetCollide
-                //      + return Adjusted.
-                // Named-retail: BSPTREE::find_collisions airborne branch → set_collide.
+                // ─── SetCollide response (shallow / walkable) ───────────
+                // Per retail (acclient_2013_pseudo_c.txt:323783-323821).
                 path.SetCollide(worldNormal0);
                 path.WalkableAllowance = PhysicsGlobals.LandingZ;
                 return TransitionState.Adjusted;
@@ -1650,8 +1660,7 @@ public static class BSPQuery
                 {
                     var worldNormal1 = L2W(hitPoly1!.Plane.Normal);
 
-                    // L.4-reject-steep-landing: same steep-poly slide
-                    // for head-sphere hits.
+                    // L.4 slide-tangent: same steep-poly slide for head-sphere.
                     if (worldNormal1.Z < PhysicsGlobals.FloorZ)
                     {
                         Vector3 currWorld = path.GlobalCurrCenter[0].Origin;
@@ -1666,7 +1675,7 @@ public static class BSPQuery
                         return TransitionState.Slid;
                     }
 
-                    // Head sphere hit shallow surface: same SetCollide response.
+                    // Head sphere hit shallow surface: SetCollide.
                     path.SetCollide(worldNormal1);
                     path.WalkableAllowance = PhysicsGlobals.LandingZ;
                     return TransitionState.Adjusted;
diff --git a/src/AcDream.Core/Physics/PhysicsEngine.cs b/src/AcDream.Core/Physics/PhysicsEngine.cs
index 2fee39b..da6711e 100644
--- a/src/AcDream.Core/Physics/PhysicsEngine.cs
+++ b/src/AcDream.Core/Physics/PhysicsEngine.cs
@@ -594,6 +594,25 @@ public sealed class PhysicsEngine
                 body.SlidingNormal = Vector3.Zero;
                 body.TransientState &= ~TransientStateFlags.Sliding;
             }
+
+            // L.4 retail-strict (2026-04-30): apply OBJECTINFO::kill_velocity.
+            // Phase 3's reset path sets VelocityKilled when an airborne hit
+            // can't find a walkable surface (steep roof, wall) AND the
+            // body had a last_known_contact_plane (i.e., was grounded
+            // recently). Retail zeros all three velocity components so
+            // gravity restarts cleanly next frame.
+            //
+            // Named-retail: OBJECTINFO::kill_velocity → CPhysicsObj::set_velocity({0,0,0}, 0)
+            //   acclient_2013_pseudo_c.txt:274467-274475
+            // Called from CTransition::transitional_insert reset path:
+            //   acclient_2013_pseudo_c.txt:273237 (Phase 3)
+            //   acclient_2013_pseudo_c.txt:272567 (validate_transition)
+            if (transition.ObjectInfo.VelocityKilled)
+            {
+                if (Environment.GetEnvironmentVariable("ACDREAM_DUMP_STEEP_ROOF") == "1")
+                    Console.WriteLine($"[steep-roof] KILL-VELOCITY-APPLIED Vbefore=({body.Velocity.X:F2},{body.Velocity.Y:F2},{body.Velocity.Z:F2}) → 0,0,0");
+                body.Velocity = Vector3.Zero;
+            }
         }
 
         // L.3a (2026-04-30): surface the wall normal so callers can apply
diff --git a/src/AcDream.Core/Physics/TransitionTypes.cs b/src/AcDream.Core/Physics/TransitionTypes.cs
index 7f28d4f..c7db51d 100644
--- a/src/AcDream.Core/Physics/TransitionTypes.cs
+++ b/src/AcDream.Core/Physics/TransitionTypes.cs
@@ -79,12 +79,31 @@ public sealed class ObjectInfo
         => OnWalkable ? PhysicsGlobals.FloorZ : PhysicsGlobals.LandingZ;
 
     /// <summary>
-    /// Stop any accumulated velocity on this object info.
-    /// ACE: ObjectInfo.StopVelocity — clears Velocity on the physics body.
-    /// acdream: velocity is tracked on PhysicsBody, not here. No-op for now;
-    /// will be wired when velocity is threaded through TransitionalInsert.
+    /// Sticky flag: set by <see cref="StopVelocity"/>; PhysicsEngine consumes
+    /// it after the transition commits to zero the body's velocity. Models
+    /// retail's <c>OBJECTINFO::kill_velocity → CPhysicsObj::set_velocity({0,0,0}, 0)</c>
+    /// (named-retail acclient_2013_pseudo_c.txt:274467-274475).
+    /// Cleared by the engine when consumed; reset to false at the start of
+    /// each <c>FindTransitionalPosition</c>.
     /// </summary>
-    public void StopVelocity() { /* velocity lives on PhysicsBody, not here */ }
+    public bool VelocityKilled;
+
+    /// <summary>
+    /// Stop any accumulated velocity on this object info.
+    /// Retail: <c>OBJECTINFO::kill_velocity</c> calls
+    /// <c>CPhysicsObj::set_velocity(object, {0,0,0}, 0)</c>
+    /// (named-retail 0x50cfe0).
+    /// ACE: <c>ObjectInfo.StopVelocity</c> — clears Velocity on the physics body.
+    /// <para>
+    /// Velocity lives on <see cref="PhysicsBody"/>, not here. We can't reach
+    /// the body directly from inside the resolver without coupling
+    /// ObjectInfo to it, so we set a flag and let
+    /// <see cref="PhysicsEngine.ResolveWithTransition"/> apply the zero
+    /// after the transition completes. The flag is sticky across the
+    /// outer step loop and consumed exactly once per resolve.
+    /// </para>
+    /// </summary>
+    public void StopVelocity() { VelocityKilled = true; }
 }
 
 /// <summary>
@@ -439,6 +458,13 @@ public sealed class Transition
     {
         var sp = SpherePath;
 
+        // L.4 retail-strict (2026-04-30): clear the kill_velocity flag at
+        // the start of each resolve so leftover state from a prior
+        // transition doesn't carry over. Inside the loop, Phase 3's reset
+        // path may set this via OBJECTINFO::StopVelocity; the engine reads
+        // it after FindTransitionalPosition returns.
+        ObjectInfo.VelocityKilled = false;
+
         // No starting cell → cannot move.
         if (sp.CurCellId == 0)
             return false;
@@ -676,13 +702,50 @@ public sealed class Transition
                     ci.ContactPlaneValid   = false;
                     ci.ContactPlaneIsWater = false;
 
+                    bool diagSteep = Environment.GetEnvironmentVariable("ACDREAM_DUMP_STEEP_ROOF") == "1";
+                    if (diagSteep)
+                    {
+                        Console.WriteLine(
+                            $"[steep-roof] PHASE3-RESET lastKnownValid={ci.LastKnownContactPlaneValid} " +
+                            $"checkPos=({sp.CheckPos.X:F2},{sp.CheckPos.Y:F2},{sp.CheckPos.Z:F2}) " +
+                            $"curPos=({sp.CurPos.X:F2},{sp.CurPos.Y:F2},{sp.CurPos.Z:F2}) " +
+                            $"stepUpNormal=({sp.StepUpNormal.X:F2},{sp.StepUpNormal.Y:F2},{sp.StepUpNormal.Z:F2})");
+                    }
+
+                    // Retail-faithful gate (acclient_2013_pseudo_c.txt:273231-273239):
+                    //
+                    //   if (last_known_valid == 0) {
+                    //       set_collision_normal(step_up_normal); return COLLIDED;
+                    //   }
+                    //   kill_velocity(this);
+                    //   last_known_valid = 0;
+                    //   return COLLIDED;
+                    //
+                    // kill_velocity ONLY fires when last_known was valid. When
+                    // it's not (the case our L.2.4 proximity guard produces
+                    // after a few airborne frames), velocity is PRESERVED so
+                    // the bounce reflection in handle_all_collisions can
+                    // redirect V's perpendicular component along the slope's
+                    // tangent direction — that's how retail's body escapes
+                    // the wedge geometry.
+                    //
+                    // This was deviated to "unconditional" earlier in this
+                    // session as a hypothesis-driven fix; the live trace
+                    // showed the deviation CAUSED the wedge by zeroing V
+                    // every frame, leaving the body with no tangent momentum
+                    // to escape (live diag 2026-04-30: V=(0,0,0) for 169
+                    // consecutive frames while position pre/resolved frozen).
                     if (ci.LastKnownContactPlaneValid)
                     {
                         ci.LastKnownContactPlaneValid = false;
                         oi.StopVelocity();
+                        if (diagSteep)
+                            Console.WriteLine($"[steep-roof] PHASE3-RESET-KILLV ← StopVelocity called");
                     }
                     else
+                    {
                         ci.SetCollisionNormal(sp.StepUpNormal);
+                    }
 
                     return TransitionState.Collided;
                 }
diff --git a/tools/pdb-extract/check_exe_pdb.py b/tools/pdb-extract/check_exe_pdb.py
new file mode 100644
index 0000000..e2e27f6
--- /dev/null
+++ b/tools/pdb-extract/check_exe_pdb.py
@@ -0,0 +1,119 @@
+"""Check an .exe's CodeView debug info to see what PDB GUID + age it
+expects. Used to verify whether a candidate acclient.exe matches our
+acclient.pdb without running the binary.
+
+Usage:
+    py tools/pdb-extract/check_exe_pdb.py <path-to-exe>
+"""
+
+import struct
+import sys
+import uuid
+
+
+def main():
+    if len(sys.argv) < 2:
+        print("usage: check_exe_pdb.py <path-to-exe>")
+        sys.exit(1)
+
+    with open(sys.argv[1], "rb") as f:
+        data = f.read()
+
+    # DOS header -> e_lfanew @ offset 0x3C points to PE header
+    pe_off = struct.unpack_from("<I", data, 0x3C)[0]
+    assert data[pe_off:pe_off + 4] == b"PE\x00\x00", "not a PE file"
+
+    # COFF header
+    machine = struct.unpack_from("<H", data, pe_off + 4)[0]
+    n_sections = struct.unpack_from("<H", data, pe_off + 6)[0]
+    timestamp = struct.unpack_from("<I", data, pe_off + 8)[0]
+    opt_size = struct.unpack_from("<H", data, pe_off + 20)[0]
+
+    print(f"machine        = 0x{machine:04x}")
+    print(f"timestamp      = 0x{timestamp:08x} ({timestamp})")
+
+    import datetime
+    ts = datetime.datetime.fromtimestamp(timestamp, tz=datetime.timezone.utc)
+    print(f"  -> linker UTC: {ts.isoformat()}")
+
+    # Optional header — magic indicates 32 vs 64 bit
+    opt_off = pe_off + 24
+    magic = struct.unpack_from("<H", data, opt_off)[0]
+    is_pe32_plus = (magic == 0x20B)
+    print(f"opt magic      = 0x{magic:04x} ({'PE32+' if is_pe32_plus else 'PE32'})")
+
+    # Data directories: PE32 has them at opt_off + 96; PE32+ at opt_off + 112
+    dd_off = opt_off + (112 if is_pe32_plus else 96)
+    # Debug directory is data dir [6]
+    debug_va = struct.unpack_from("<I", data, dd_off + 6 * 8)[0]
+    debug_size = struct.unpack_from("<I", data, dd_off + 6 * 8 + 4)[0]
+    print(f"debug dir      = VA=0x{debug_va:08x}  size={debug_size}")
+
+    # We need to map the VA back to a file offset via section headers
+    sec_off = opt_off + opt_size
+    sections = []
+    for i in range(n_sections):
+        s = sec_off + i * 40
+        name = data[s:s + 8].rstrip(b"\x00").decode("ascii", errors="replace")
+        vsize = struct.unpack_from("<I", data, s + 8)[0]
+        vaddr = struct.unpack_from("<I", data, s + 12)[0]
+        rsize = struct.unpack_from("<I", data, s + 16)[0]
+        roff = struct.unpack_from("<I", data, s + 20)[0]
+        sections.append((name, vaddr, vsize, roff, rsize))
+
+    def va_to_file(va):
+        for (name, vaddr, vsize, roff, rsize) in sections:
+            if vaddr <= va < vaddr + vsize:
+                return roff + (va - vaddr)
+        return None
+
+    debug_off = va_to_file(debug_va)
+    if debug_off is None:
+        print("debug directory VA does not map into any section")
+        return
+
+    # Each debug directory entry is 28 bytes
+    n_entries = debug_size // 28
+    print(f"# debug entries = {n_entries}")
+
+    for i in range(n_entries):
+        e = debug_off + i * 28
+        characteristics = struct.unpack_from("<I", data, e)[0]
+        ts_e = struct.unpack_from("<I", data, e + 4)[0]
+        major = struct.unpack_from("<H", data, e + 8)[0]
+        minor = struct.unpack_from("<H", data, e + 10)[0]
+        type_e = struct.unpack_from("<I", data, e + 12)[0]
+        sz = struct.unpack_from("<I", data, e + 16)[0]
+        rva = struct.unpack_from("<I", data, e + 20)[0]
+        ptr = struct.unpack_from("<I", data, e + 24)[0]
+
+        type_name = {2: "CODEVIEW", 4: "MISC", 12: "VC_FEATURE", 13: "POGO", 16: "REPRO"}.get(type_e, f"type_{type_e}")
+        print(f"  entry {i}: type={type_name} sz={sz} fileOff=0x{ptr:08x}")
+
+        if type_e == 2 and sz >= 24:
+            cv = data[ptr:ptr + sz]
+            sig = cv[:4]
+            print(f"    cv signature = {sig!r}")
+            if sig == b"RSDS":
+                guid_bytes = cv[4:20]
+                age = struct.unpack_from("<I", cv, 20)[0]
+                pdb_name = cv[24:].rstrip(b"\x00").decode("utf-8", errors="replace")
+                pdb_guid = uuid.UUID(bytes_le=guid_bytes)
+                print(f"    GUID         = {{{pdb_guid}}}")
+                print(f"    age          = {age}")
+                print(f"    PDB filename = {pdb_name}")
+
+                expected_guid = uuid.UUID("9e847e2f-777c-4bd9-886c-22256bb87f32")
+                expected_age = 1
+                if pdb_guid == expected_guid and age == expected_age:
+                    print()
+                    print("=== MATCH: this exe pairs with our acclient.pdb ===")
+                else:
+                    print()
+                    print("=== MISMATCH ===")
+                    print(f"  expected GUID = {{{expected_guid}}}")
+                    print(f"  expected age  = {expected_age}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/tools/pdb-extract/dump_pdb_info.py b/tools/pdb-extract/dump_pdb_info.py
new file mode 100644
index 0000000..bac3076
--- /dev/null
+++ b/tools/pdb-extract/dump_pdb_info.py
@@ -0,0 +1,98 @@
+"""Dump the PDB info stream so we know exactly which acclient.exe build
+matches our PDB GUID. The PDB header points to stream 1 ("PDB Info") which
+contains: u32 version, u32 signature(timestamp), u32 age, 16-byte GUID.
+
+Usage:
+    py tools/pdb-extract/dump_pdb_info.py refs/acclient.pdb
+"""
+
+import struct
+import sys
+import datetime
+import uuid
+
+
+def _ceil_div(a, b):
+    return (a + b - 1) // b
+
+
+def main():
+    if len(sys.argv) < 2:
+        print("usage: dump_pdb_info.py <path-to-pdb>")
+        sys.exit(1)
+
+    pdb_path = sys.argv[1]
+    with open(pdb_path, "rb") as f:
+        data = f.read()
+
+    magic = b"Microsoft C/C++ MSF 7.00\r\n\x1aDS\x00\x00\x00"
+    assert data.startswith(magic), "not an MSF 7.00 PDB"
+
+    block_size = struct.unpack_from("<I", data, 0x20)[0]
+    num_blocks = struct.unpack_from("<I", data, 0x28)[0]
+    num_dir_bytes = struct.unpack_from("<I", data, 0x2C)[0]
+    block_map_addr = struct.unpack_from("<I", data, 0x34)[0]
+
+    print(f"block_size      = {block_size}")
+    print(f"num_blocks      = {num_blocks}")
+    print(f"num_dir_bytes   = {num_dir_bytes}")
+    print(f"block_map_addr  = {block_map_addr}")
+
+    def read_page(idx):
+        return data[idx * block_size : (idx + 1) * block_size]
+
+    dir_pages_needed = _ceil_div(num_dir_bytes, block_size)
+    block_map = read_page(block_map_addr)
+    dir_page_indices = struct.unpack_from(f"<{dir_pages_needed}I", block_map, 0)
+    dir_data = bytearray()
+    for pi in dir_page_indices:
+        dir_data.extend(read_page(pi))
+    dir_data = bytes(dir_data)
+
+    num_streams = struct.unpack_from("<I", dir_data, 0)[0]
+    stream_sizes = struct.unpack_from(f"<{num_streams}I", dir_data, 4)
+    print(f"num_streams     = {num_streams}")
+
+    offset = 4 + num_streams * 4
+    streams = []
+    for sz in stream_sizes:
+        if sz == 0xFFFFFFFF:
+            streams.append((0, []))
+            continue
+        n_pages = _ceil_div(sz, block_size)
+        pages = struct.unpack_from(f"<{n_pages}I", dir_data, offset)
+        offset += n_pages * 4
+        streams.append((sz, list(pages)))
+
+    # Stream 1 = PDB Info Stream
+    pdb_info_size, pdb_info_pages = streams[1]
+    print(f"pdb_info_size   = {pdb_info_size}")
+
+    pdb_info = bytearray()
+    for pi in pdb_info_pages:
+        pdb_info.extend(read_page(pi))
+    pdb_info = bytes(pdb_info[:pdb_info_size])
+
+    version = struct.unpack_from("<I", pdb_info, 0)[0]
+    signature = struct.unpack_from("<I", pdb_info, 4)[0]
+    age = struct.unpack_from("<I", pdb_info, 8)[0]
+    guid_bytes = pdb_info[12:28]
+    pdb_guid = uuid.UUID(bytes_le=guid_bytes)
+
+    sig_dt = datetime.datetime.fromtimestamp(signature, tz=datetime.timezone.utc)
+
+    print()
+    print("=== PDB Info Stream ===")
+    print(f"version    = {version}")
+    print(f"signature  = 0x{signature:08x}  ({signature})")
+    print(f"  -> linker timestamp UTC: {sig_dt.isoformat()}")
+    print(f"age        = {age}")
+    print(f"GUID       = {{{pdb_guid}}}")
+    print()
+    print("This is the GUID + age the matching acclient.exe must reference")
+    print("in its CodeView entry. Find a binary whose linker timestamp")
+    print(f"is around {sig_dt.strftime('%Y-%m-%d')}.")
+
+
+if __name__ == "__main__":
+    main()