From 8a82cebae6f9cdfdcdc01708a7a15d67e259fc81 Mon Sep 17 00:00:00 2001
From: Sam <samuv.eth@gmail.com>
Date: Thu, 20 Aug 2026 19:09:00 +0700
Subject: [PATCH] the walking tax banded, and the minimap stops rewriting the
 world
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- a panning UI batch rides its Transform instead of rewriting its mesh, minimap tiles emit unclipped (the mask shader already zeroes the corners), and the blip layer's origin snaps to half a logical pixel — steady UI mesh writes while moving drop from ~183/s to 10–14/s, and three minimap quads stop arming change detection over ~40k world meshes
- the classify twin GC comes off the despawn edge: a part despawn schedules the full walk behind a 2 s deadline instead of running it that frame — 12–22 full walks a second while walking, gone
- the walking tax measured and banded at Goldshire: parked 10.7 → moving 14.1 ms/frame of process CPU, ~+2.5 after the two fixes above
- WOW_PART_CHURN and WOW_MESH_HOLDERS count part spawn/despawn frames and name which entities hold a churning mesh; the capture rig's `at:` pin learns a facing

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 crates/benilla-app/src/capture/probe_rig.rs |  27 ++++-
 crates/benilla-app/src/minimap/mod.rs       |  18 ++-
 crates/benilla-app/src/perf/census.rs       |  95 +++++++++++++++
 crates/benilla-app/src/perf/mod.rs          |  11 ++
 crates/benilla-app/src/ui_pass.rs           | 121 +++++++++++++++++++-
 crates/benilla-world/src/model_render.rs    |  40 +++++--
 6 files changed, 289 insertions(+), 23 deletions(-)

diff --git a/crates/benilla-app/src/capture/probe_rig.rs b/crates/benilla-app/src/capture/probe_rig.rs
index be7078a8..a31a7ce4 100644
--- a/crates/benilla-app/src/capture/probe_rig.rs
+++ b/crates/benilla-app/src/capture/probe_rig.rs
@@ -54,7 +54,7 @@
 //! | `gear:<t>` | `.character premade gear <t>` | SEC_BASIC_ADMIN 4 | name or entry; levels up + equips |
 //! | `spec:<t>` | `.character premade spec <t>` | SEC_BASIC_ADMIN 4 | resets talents, learns the tree |
 //! | `at:<name>` | `.tele <name>` | SEC_TICKETMASTER 2 | 997 named rows in `game_tele` |
-//! | `at:m,x,y,z` | `.go xyz x y z m` | SEC_TICKETMASTER 2 | |
+//! | `at:m,x,y,z[,o]` | `.go xyz x y z m` (`.go xyzo` with `o`) | SEC_TICKETMASTER 2 | pin the facing when a motion probe walks |
 //! | `gm:on\|off` | `.gm on\|off` | SEC_GAMEMASTER 3 | see 0649 on why `off` matters |
 //!
 //! Probe accounts are gmlevel **6**, so every one of these lands. Pass `gear:?` (or `spec:?`) to
@@ -506,15 +506,21 @@ fn build_steps(spec: &RigSpec, dead: bool) -> Vec<String> {
     }
     if let Some(at) = &spec.at {
         steps.push(match parse_point(at) {
-            Some((map, x, y, z)) => format!(".go xyz {x} {y} {z} {map}"),
+            Some((map, x, y, z, Some(o))) => format!(".go xyzo {x} {y} {z} {o} {map}"),
+            Some((map, x, y, z, None)) => format!(".go xyz {x} {y} {z} {map}"),
             None => format!(".tele {at}"),
         });
     }
     steps
 }
 
-/// `map,x,y,z` → the `.go xyz` argument order (`x y z map`). Anything else is a `game_tele` name.
-fn parse_point(at: &str) -> Option<(i32, f32, f32, f32)> {
+/// `map,x,y,z[,o]` → `.go xyz x y z map`, or `.go xyzo x y z o map` when a facing is given
+/// (VERIFIED vmangos `Chat.cpp:408`: `xyzo`, SEC_TICKETMASTER like `xyz`). Anything else is a
+/// `game_tele` name. The facing is load-bearing exactly when a motion probe walks from the
+/// pin: `W` follows the BODY's facing, and each probe character keeps whatever facing it last
+/// had — an unpinned facing sent probe0 and probe4 down different routes from the same point,
+/// which invalidated a cross-slot A/B before 1462's sitting caught it.
+fn parse_point(at: &str) -> Option<(i32, f32, f32, f32, Option<f32>)> {
     let mut parts = at.split(',').map(str::trim);
     let map = parts.next()?.parse().ok()?;
     let (x, y, z) = (
@@ -522,7 +528,11 @@ fn parse_point(at: &str) -> Option<(i32, f32, f32, f32)> {
         parts.next()?.parse().ok()?,
         parts.next()?.parse().ok()?,
     );
-    parts.next().is_none().then_some((map, x, y, z))
+    let o = match parts.next() {
+        Some(o) => Some(o.parse().ok()?),
+        None => None,
+    };
+    parts.next().is_none().then_some((map, x, y, z, o))
 }
 
 fn rig_char_name(spec: &RigSpec) -> Option<String> {
@@ -743,9 +753,14 @@ mod tests {
             build_steps(&point, false),
             vec![".go xyz -1277.5 124 131.2 1"]
         );
+        let faced = RigSpec::parse("at:0,-9399.03,10.41,59.83,0.0").unwrap();
+        assert_eq!(
+            build_steps(&faced, false),
+            vec![".go xyzo -9399.03 10.41 59.83 0 0"]
+        );
         assert_eq!(parse_point("ThunderBluff"), None);
         assert_eq!(parse_point("1,2,3"), None); // three numbers is not a map + point
-        assert_eq!(parse_point("1,2,3,4,5"), None);
+        assert_eq!(parse_point("1,2,3,4,5,6"), None); // and six is past even a faced point
     }
 
     #[test]
diff --git a/crates/benilla-app/src/minimap/mod.rs b/crates/benilla-app/src/minimap/mod.rs
index 24ee91a9..30240f90 100644
--- a/crates/benilla-app/src/minimap/mod.rs
+++ b/crates/benilla-app/src/minimap/mod.rs
@@ -517,7 +517,11 @@ fn emit_minimap(
                     z_key: slot.z,
                     texture: Some(handle.clone()),
                     color: [tint[0], tint[1], tint[2], slot.alpha],
-                    clip: Some(slot.rect),
+                    // No CPU clip (1463): the mask shader already zeroes everything outside
+                    // `mask_rect` (`ui_quad.wgsl`'s `inside` test), and clipping a PANNING tile
+                    // re-cut its quad every frame — constant positions, churning UVs — which is
+                    // exactly the shape the batcher's pan gate cannot ride on a `Transform`.
+                    // Unclipped, the tile is a pure translation and never rewrites its mesh.
                     mask: mask.clone(),
                     ..default()
                 });
@@ -552,6 +556,14 @@ fn emit_minimap(
         }
         let win = window.iter().next();
         let cursor = win.and_then(|w| w.cursor_position());
+        // The player's pan term, quantized to a half-logical-pixel grid (one device px at 2×):
+        // every blip offset — and the rim arrows' bearing — derives from `wx`/`wy`, so the whole
+        // blip layer steps together a few times a second instead of re-emitting sub-pixel-shifted
+        // quads every frame while walking (1463; a 0.07 px/frame slide on a 16 px icon is not a
+        // visible motion, but each slide rewrote the batch mesh and armed the world's
+        // `AssetChanged` scans). The blips' world positions stay exact — only the shared pan
+        // origin snaps.
+        let q = 0.5 / blip_px_per_yd;
         blips::BlipCtx {
             center,
             side,
@@ -559,8 +571,8 @@ fn emit_minimap(
             radius_yd: (side * 0.5) / blip_px_per_yd,
             z: slot.z,
             alpha: slot.alpha,
-            wx,
-            wy,
+            wx: (wx / q).round() * q,
+            wy: (wy / q).round() * q,
             wz: wow[2],
             cursor,
             // The same point in UI space (y-up, ÷s through the 0582/0584 seam — the tooltip's
diff --git a/crates/benilla-app/src/perf/census.rs b/crates/benilla-app/src/perf/census.rs
index d1ee902d..737b7cba 100644
--- a/crates/benilla-app/src/perf/census.rs
+++ b/crates/benilla-app/src/perf/census.rs
@@ -295,6 +295,101 @@ pub(super) fn count_mesh_events(
     }
 }
 
+/// `WOW_PART_CHURN=1`: per-second M2-part churn. `rm_frames` is the count of frames with ≥1
+/// `MeshMaterial3d<WowModelMaterial>` removal — exactly the predicate that promotes
+/// `classify_water_side` to its full walk (0930's twin-GC mark), so a regime reading ~60 here
+/// full-walks every frame. `mat_mod` counts `WowModelMaterial` asset Modified events — the
+/// global asset-changed tick that wakes every `AssetChanged` scan.
+pub(super) fn count_part_churn(
+    added: Query<(), Added<MeshMaterial3d<benilla_assets::materials::WowModelMaterial>>>,
+    mut removed: RemovedComponents<MeshMaterial3d<benilla_assets::materials::WowModelMaterial>>,
+    mut mat_events: MessageReader<
+        bevy::asset::AssetEvent<benilla_assets::materials::WowModelMaterial>,
+    >,
+    time: Res<Time>,
+    mut acc: Local<(f32, u32, u32, u32, u32)>,
+) {
+    let (last, add_n, rm_n, rm_frames, mat_mod) = &mut *acc;
+    *add_n += added.iter().count() as u32;
+    let rm = removed.read().count() as u32;
+    *rm_n += rm;
+    if rm > 0 {
+        *rm_frames += 1;
+    }
+    *mat_mod += mat_events
+        .read()
+        .filter(|e| matches!(e, bevy::asset::AssetEvent::Modified { .. }))
+        .count() as u32;
+    if time.elapsed_secs() - *last >= 1.0 {
+        eprintln!(
+            "[part-churn] parts +{add_n} -{rm_n} rm_frames={rm_frames}/s mat_modified={mat_mod}/s"
+        );
+        (*add_n, *rm_n, *rm_frames, *mat_mod) = (0, 0, 0, 0);
+        *last = time.elapsed_secs();
+    }
+}
+
+/// `WOW_MESH_HOLDERS=1`: once a second, the set of mesh asset ids Modified in the last second
+/// and the component signature of every entity holding one — [`count_mesh_events`] counts and
+/// samples the ids, this names the writer by its holder's archetype (an id alone says nothing).
+/// Exclusive, like [`arch_census`], for the same reason: the signature needs the live archetype.
+pub(super) fn mesh_holders(
+    world: &mut World,
+    mut cursor: Local<Option<bevy::ecs::message::MessageCursor<bevy::asset::AssetEvent<Mesh>>>>,
+    mut acc: Local<(f32, std::collections::HashSet<bevy::asset::AssetId<Mesh>>)>,
+) {
+    let messages = world.resource::<bevy::ecs::message::Messages<bevy::asset::AssetEvent<Mesh>>>();
+    let cursor = cursor.get_or_insert_with(|| messages.get_cursor());
+    let (last, ids) = &mut *acc;
+    ids.extend(cursor.read(messages).filter_map(|e| match e {
+        bevy::asset::AssetEvent::Modified { id } => Some(*id),
+        _ => None,
+    }));
+    let now = world.resource::<Time>().elapsed_secs();
+    if now - *last < 1.0 {
+        return;
+    }
+    *last = now;
+    if ids.is_empty() {
+        return;
+    }
+    let short = |full: &str| -> String {
+        let base = full.split('<').next().unwrap_or(full);
+        let segs: Vec<&str> = base.split("::").collect();
+        segs[segs.len().saturating_sub(2)..].join("::")
+    };
+    let mut found = 0usize;
+    let mut holders = world.query::<(Entity, &Mesh3d)>();
+    let matches: Vec<Entity> = holders
+        .iter(world)
+        .filter(|(_, m)| ids.contains(&m.0.id()))
+        .map(|(e, _)| e)
+        .take(6)
+        .collect();
+    for e in matches {
+        found += 1;
+        let sig: Vec<String> = world
+            .entity(e)
+            .archetype()
+            .components()
+            .iter()
+            .filter_map(|&c| world.components().get_info(c))
+            .map(|i| short(&i.name().to_string()))
+            .collect();
+        eprintln!(
+            "[mesh-holders] {e} holds a modified mesh: {}",
+            sig.join("+")
+        );
+    }
+    if found == 0 {
+        eprintln!(
+            "[mesh-holders] {} modified id(s), no Mesh3d holder (2d or unowned)",
+            ids.len()
+        );
+    }
+    ids.clear();
+}
+
 /// When the one-shot archetype census fires (seconds of `Time` elapsed; `f32::MAX` = spent).
 #[derive(Resource)]
 pub(super) struct ArchCensusAt(pub(super) f32);
diff --git a/crates/benilla-app/src/perf/mod.rs b/crates/benilla-app/src/perf/mod.rs
index 7eb285ba..88b72771 100644
--- a/crates/benilla-app/src/perf/mod.rs
+++ b/crates/benilla-app/src/perf/mod.rs
@@ -99,6 +99,17 @@ impl Plugin for PerfPlugin {
         if std::env::var_os("WOW_MESH_EVENTS").is_some() {
             app.add_systems(Update, census::count_mesh_events);
         }
+        // `WOW_PART_CHURN=1` — the moving-regime premise counters (1461): how many frames per
+        // second despawn an M2 part (= `classify_water_side` full-walk promotions), and how
+        // often model materials fire Modified (= the `AssetChanged` scan wake-ups).
+        if std::env::var_os("WOW_PART_CHURN").is_some() {
+            app.add_systems(Update, census::count_part_churn);
+        }
+        // `WOW_MESH_HOLDERS=1` — who OWNS the meshes `WOW_MESH_EVENTS` counts? Prints each
+        // holder's archetype signature once a second (1461's writer hunt).
+        if std::env::var_os("WOW_MESH_HOLDERS").is_some() {
+            app.add_systems(Last, census::mesh_holders);
+        }
         // `WOW_CAM_CHANGED=1` — is the world camera's transform bit-stable at a parked pin?
         // The premise counter for gating the per-submesh visibility sweep on frame-stable
         // camera inputs: if the controller rewrites an equal transform every frame, change
diff --git a/crates/benilla-app/src/ui_pass.rs b/crates/benilla-app/src/ui_pass.rs
index 58d433d1..de736ddb 100644
--- a/crates/benilla-app/src/ui_pass.rs
+++ b/crates/benilla-app/src/ui_pass.rs
@@ -656,10 +656,18 @@ impl Run {
 struct BatchPools {
     entities: Vec<Entity>,
     meshes: Vec<Handle<Mesh>>,
-    /// Each slot's last-written content — the per-run skip gate (decision 1361): a slot whose
-    /// run is bit-identical to what its pooled mesh already holds is not rewritten, so one
-    /// animating quad no longer drags every batch through the GPU mesh allocator.
+    /// Each slot's last-written MESH content (base positions, before [`Self::offsets`]) — the
+    /// per-run skip gate (decision 1361): a slot whose run is bit-identical to what its pooled
+    /// mesh already holds is not rewritten, so one animating quad no longer drags every batch
+    /// through the GPU mesh allocator.
     stored: Vec<StoredRun>,
+    /// Each slot's current XY translation from its stored base, carried by the batch entity's
+    /// `Transform` instead of the mesh (decision 1463): a run that only *panned* — the minimap
+    /// tile was ~74% of all moving-regime rebuild triggers — moves without an `Assets<Mesh>`
+    /// write, because ONE Modified event per frame arms `AssetChanged` probes over every
+    /// `Mesh3d` row in the scene (1370's all-or-nothing fast path, ~0.8 ms/frame at 1461's
+    /// Goldshire pin).
+    offsets: Vec<Vec2>,
     materials: std::collections::HashMap<MatKey, Handle<UiQuadMaterial>>,
 }
 
@@ -676,6 +684,55 @@ struct StoredRun {
     key: MatKey,
 }
 
+impl StoredRun {
+    /// The candidate's uniform XY offset from this base: `Some(d)` when every vertex is this
+    /// run's vertex plus one constant delta and everything else is bit-equal — `Some(ZERO)` is
+    /// exactly the old bit-identical skip case, so this subsumes 1361's gate. `None` means the
+    /// run genuinely changed shape and must be rebaked. The comparison is exact float
+    /// arithmetic on purpose: a miss only falls back to the rewrite path, so a widget whose
+    /// pan doesn't survive `b + d == n` bit-exactly loses the optimization, never correctness.
+    /// Why [`Self::translation_from`] missed — the pan-gate diagnostic (`WOW_UI_DIFF=1`).
+    fn translation_miss_reason(&self, new: &StoredRun) -> &'static str {
+        if self.key != new.key {
+            "key"
+        } else if self.z_bits != new.z_bits {
+            "z"
+        } else if self.positions.len() != new.positions.len() {
+            "len"
+        } else if self.indices != new.indices {
+            "indices"
+        } else if self.uvs != new.uvs {
+            "uvs"
+        } else if self.colors != new.colors {
+            "colors"
+        } else {
+            "delta-nonuniform"
+        }
+    }
+
+    fn translation_from(&self, new: &StoredRun) -> Option<Vec2> {
+        if self.key != new.key
+            || self.z_bits != new.z_bits
+            || self.positions.len() != new.positions.len()
+            || self.positions.is_empty()
+            || self.indices != new.indices
+            || self.uvs != new.uvs
+            || self.colors != new.colors
+        {
+            return None;
+        }
+        let d = Vec2::new(
+            new.positions[0][0] - self.positions[0][0],
+            new.positions[0][1] - self.positions[0][1],
+        );
+        self.positions
+            .iter()
+            .zip(&new.positions)
+            .all(|(b, n)| n[0] == b[0] + d.x && n[1] == b[1] + d.y && n[2] == b[2])
+            .then_some(d)
+    }
+}
+
 /// A material's full identity: texture, blend/shape flags, mask texture + mask rect (as bits, so
 /// NaN/-0.0 can't split cache entries byte-equal materials would share).
 type MatKey = (
@@ -696,8 +753,10 @@ fn retire_batches(pools: &mut BatchPools, commands: &mut Commands) {
         commands.entity(entity).despawn();
     }
     // The skip gate must forget with them (decision 1361): a retired slot's entity is gone, so
-    // "content unchanged" must not skip the respawn when the UI returns.
+    // "content unchanged" must not skip the respawn when the UI returns. The pan gate's
+    // offsets ride the same lifecycle (they index the same slots).
     pools.stored.clear();
+    pools.offsets.clear();
 }
 
 /// Per frame: when the BASE lane flagged a change ([`UiQuads::dirty`]) or the APPEND lane's
@@ -962,10 +1021,54 @@ fn rebuild_ui_mesh(
             z_bits: z.to_bits(),
             key,
         };
-        if pools.stored.get(used) == Some(&stored) {
+        // The pan gate (1463): a run that matches its slot's base up to one constant XY delta
+        // moves on the batch entity's `Transform` — no mesh write, no `AssetChanged` arming.
+        // `Some(ZERO)` is the bit-identical case (1361's old skip gate), where even the
+        // `Transform` write is skipped unless a previous pan is being undone.
+        let pan = pools
+            .stored
+            .get(used)
+            .and_then(|prev| prev.translation_from(&stored).map(|d| (d, prev.z_bits)));
+        if let Some((d, z_bits)) = pan {
+            if pools.offsets[used] != d {
+                pools.offsets[used] = d;
+                if let Some(&entity) = pools.entities.get(used) {
+                    commands.entity(entity).insert(Transform::from_xyz(
+                        d.x,
+                        d.y,
+                        f32::from_bits(z_bits),
+                    ));
+                }
+            }
             used += 1;
             continue;
         }
+        // Pan-gate miss diagnostic (`WOW_UI_DIFF=1`, ≤3 lines/s so a startup burst can't
+        // exhaust it): names the check that sent this slot to the rewrite path.
+        if *UI_DIFF {
+            use std::sync::atomic::{AtomicU32, AtomicU64, Ordering};
+            static SHOWN: AtomicU32 = AtomicU32::new(0);
+            static LAST_SEC: AtomicU64 = AtomicU64::new(0);
+            static START: std::sync::OnceLock<std::time::Instant> = std::sync::OnceLock::new();
+            let sec = START
+                .get_or_init(std::time::Instant::now)
+                .elapsed()
+                .as_secs();
+            if LAST_SEC.swap(sec, Ordering::Relaxed) != sec {
+                SHOWN.store(0, Ordering::Relaxed);
+            }
+            if SHOWN.fetch_add(1, Ordering::Relaxed) < 3 {
+                let why = pools
+                    .stored
+                    .get(used)
+                    .map_or("no-prev", |p| p.translation_miss_reason(&stored));
+                eprintln!(
+                    "[ui-pan] slot {used} rewrite: {why} (quads={}, key.tex={:?})",
+                    stored.positions.len() / 4,
+                    stored.key.0
+                );
+            }
+        }
         // MAIN_WORLD too (not RENDER_WORLD-only): the pool rewrites this asset in place next
         // rebuild, so the main-world copy must survive extraction.
         let mut mesh = Mesh::new(
@@ -981,6 +1084,13 @@ fn rebuild_ui_mesh(
         } else {
             pools.stored.push(stored);
         }
+        // A rebaked slot's mesh holds absolute positions again — its pan resets with it (the
+        // rewrite arm below writes `Transform::from_xyz(0, 0, z)`).
+        if pools.offsets.len() > used {
+            pools.offsets[used] = Vec2::ZERO;
+        } else {
+            pools.offsets.push(Vec2::ZERO);
+        }
         let mesh_handle = match pools.meshes.get(used) {
             Some(handle) => {
                 // `WOW_FREEZE_UI_MESH=1` — never rewrite a pooled batch mesh in place (the UI
@@ -1048,6 +1158,7 @@ fn rebuild_ui_mesh(
     }
     pools.meshes.truncate(used);
     pools.stored.truncate(used);
+    pools.offsets.truncate(used);
 }
 
 /// Deliberately-overlapping synthetic content proving the sort: 5 z strata × 40 quads each, offset both
diff --git a/crates/benilla-world/src/model_render.rs b/crates/benilla-world/src/model_render.rs
index eea759ef..9f5b79b2 100644
--- a/crates/benilla-world/src/model_render.rs
+++ b/crates/benilla-world/src/model_render.rs
@@ -787,12 +787,22 @@ pub struct FarSideOfWater;
 /// surfaces, and the eye's side — so a *clean* batch is skipped, and only the changed set
 /// re-classifies (~2.6k of ~40k parts on a measured Stormwind frame; the full sweep cost 2.2 ms).
 /// The frame promotes to the full walk — byte-identical to the old per-frame behaviour — whenever
-/// a global term moves: the eye crosses the surface (every verdict inverts), surfaces stream
-/// in/out, or any part despawned (the twin GC below needs its full mark; between full frames a
-/// pair orphaned by a handle overwrite — a fade settling to its cutout — lingers harmlessly, kept
-/// fresh by the mirror above, until the next edge sweeps it). The room-claim trigger fans DOWN:
-/// the claim lives on the unit root, the batches are its descendants, and a claim can change with
-/// the unit standing still (a building streaming in under it re-claims at the same spot).
+/// a global term that changes VERDICTS moves: the eye crosses the surface (every verdict
+/// inverts) or surfaces stream in/out. A part *despawn* changes no live verdict — it only
+/// obligates the twin GC's exact mark eventually — so it schedules a full walk behind
+/// [`GC_DEADLINE_SECS`] instead of promoting the same frame (decision 1462: walking streams
+/// parts out on 12–22 frames/s, and each promotion was a full re-classify of every part —
+/// lane B of 1461's walking tax; between full frames a pair orphaned by a handle overwrite —
+/// a fade settling to its cutout — lingers harmlessly, kept fresh by the mirror above, until
+/// the next edge sweeps it, and a GC-due twin lingers the same way). The room-claim trigger
+/// fans DOWN: the claim lives on the unit root, the batches are its descendants, and a claim
+/// can change with the unit standing still (a building streaming in under it re-claims at the
+/// same spot).
+/// How long a despawn-obligated twin GC may wait for its full walk (1462). Bounds how long an
+/// orphaned twin pair can outlive its last user — a handful of small material clones — while a
+/// walking regime's continuous despawn stream costs ~0.5 full walks/s instead of 12–22.
+const GC_DEADLINE_SECS: f32 = 2.0;
+
 #[allow(clippy::type_complexity, clippy::too_many_arguments)]
 pub(crate) fn classify_water_side(
     interleave: crate::particles::WaterInterleave,
@@ -834,7 +844,9 @@ pub(crate) fn classify_water_side(
     reclaimed: Query<Entity, Changed<crate::wmo_portal::UnitWmoRoom>>,
     children: Query<&Children>,
     mut removed: RemovedComponents<MeshMaterial3d<WowModelMaterial>>,
+    time: Res<bevy::time::Time>,
     mut eye_was_submerged: Local<Option<bool>>,
+    mut gc_due: Local<Option<f32>>,
 ) {
     // Mirror any near-asset edit into the live twin. Since decision 1381 the texanim/tint lane
     // animates through the shared table (the twin's clone carries the same slot and seed, so it
@@ -858,12 +870,22 @@ pub(crate) fn classify_water_side(
         }
     }
     let eye = interleave.eye_submerged();
-    // `last()` drains the reader — `next()` would leave the tail readable and promote NEXT frame
-    // to a spurious second full walk.
+    let now = time.elapsed_secs();
+    // `last()` drains the reader — `next()` would leave the tail readable and re-arm NEXT frame
+    // as a spurious second deadline. The earliest despawn holds the deadline (`get_or_insert`):
+    // a steady despawn stream must not push the GC out forever.
+    if removed.read().last().is_some() {
+        gc_due.get_or_insert(now + GC_DEADLINE_SECS);
+    }
     let full = *eye_was_submerged != Some(eye)
         || interleave.surfaces_changed()
-        || removed.read().last().is_some();
+        || gc_due.is_some_and(|due| now >= due);
     *eye_was_submerged = Some(eye);
+    if full {
+        // Any full walk runs the GC's exact mark, whatever triggered it — a pending deadline
+        // is satisfied, not merely postponed.
+        *gc_due = None;
+    }
 
     if full {
         // The full walk — every part, plus the twin GC's exact mark-and-sweep.