/target, /assist and /follow, and right click opens a clam

- /target and /assist find units by name like the original, and /follow walks you behind your target - right click opens clams and lockboxes, and the tooltip promises it - the freeze thaws clean, poses release instead of sticking, and you drop to neutral standing before it grips - unit frame art layers over the bars like the original Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-08-02 18:02:44 +07:00
parent 2f40b6a9a0
commit 94de741cb1
36 changed files with 2086 additions and 204 deletions
@@ -22,7 +22,7 @@ use benilla_protocol::{decode, EntityKind, WorldSession, WORLD_PORT};
 use clap::Parser;

 use probes::{
-    Attack, Aura, Charge, Ctx, Death, EquipPackSlot, GiverStatus, GroundFx, Loot, Probe,
+    Attack, Aura, Charge, Ctx, Death, EquipPackSlot, GiverStatus, GroundFx, Loot, OpenItem, Probe,
    QueryNames, Quest, QuestItem, QuestLog, Speed, Spells, Spirit, SwapPackSlots, UsePackSlot,
    Vendor, WorldState,
 };
@@ -81,6 +81,14 @@ struct Cli {
    /// delta or a destroy on that item's guid (consumed), or an explicit cast-result refusal.
    #[arg(long)]
    use_pack_slot: Option<u8>,
+    /// Live-verify `CMSG_OPEN_ITEM`: `.additem` a Small Barnacled Clam (entry 7973 — LOOTABLE,
+    /// LockID 0, the director's own case), find it in the backpack, and run the fork a bag
+    /// right-click makes for an *openable* item — `CMSG_OPEN_ITEM(bagIndex, slot)`, never
+    /// `CMSG_USE_ITEM` (a clam has no on-use spell, so the use goes nowhere) — requiring
+    /// `SMSG_LOOT_RESPONSE` on the **item's own guid**. Releases the window and subtracts the copy
+    /// afterwards. Needs a GM account (the deploy's probes are gmlevel 6).
+    #[arg(long)]
+    open_item: bool,
    /// Send a `/say` line right after entering the world — the GM dot-command channel
    /// (`.additem 3732`, `.repairitems`, …) for setting up probe scenarios. Repeatable; the lines go
    /// out in order (e.g. `--say ".modify money 5000000" --say ".go creature 1"`).
@@ -344,6 +352,9 @@ fn main() -> Result<()> {
    if let Some(n) = cli.use_pack_slot {
        probes.push(Box::new(UsePackSlot { n }));
    }
+    if cli.open_item {
+        probes.push(Box::new(OpenItem));
+    }
    if let Some(n) = cli.equip_pack_slot {
        probes.push(Box::new(EquipPackSlot { n }));
    }
@@ -67,6 +67,7 @@ mod equip_pack_slot;
 mod giverstatus;
 mod groundfx;
 mod loot;
+mod open_item;
 mod query_names;
 mod quest;
 mod quest_item;
@@ -87,6 +88,7 @@ pub(crate) use equip_pack_slot::EquipPackSlot;
 pub(crate) use giverstatus::GiverStatus;
 pub(crate) use groundfx::GroundFx;
 pub(crate) use loot::Loot;
+pub(crate) use open_item::OpenItem;
 pub(crate) use query_names::QueryNames;
 pub(crate) use quest::Quest;
 pub(crate) use quest_item::QuestItem;
@@ -0,0 +1,122 @@
+//! `--open-item`: the openable-item wire — the fork a bag right-click makes when the clicked
+//! item's template carries the LOOTABLE bit (`ItemInfo::openable`). The client does NOT send
+//! `CMSG_USE_ITEM` for such an item (a clam has no on-use spell at all, so the use goes nowhere —
+//! the director's "there is no way to open clams"): it sends `CMSG_OPEN_ITEM(bagIndex, slot)`, and
+//! the server answers `SMSG_LOOT_RESPONSE` **on the item's own guid**, i.e. a loot window over a
+//! thing in your bag rather than a corpse in the world.
+//!
+//! This probe proves the whole round trip against the live server: add a clam, open it by bag
+//! position, require a loot response carrying rows on that guid, then release the window and
+//! subtract the copy so the character is left as found.
+
+use std::time::{Duration, Instant};
+
+use anyhow::{Context, Result};
+use benilla_protocol::{decode, SessionEvent};
+
+use crate::probes::{Ctx, Probe};
+
+/// The probe target: "Small Barnacled Clam" (entry 7973) — the director's own case. Template
+/// `Flags` carries LOOTABLE (`0x4`) with `LockID = 0`, so it is openable the moment it exists: no
+/// key, no lockpicking, no rigging. Its `item_loot_template` rows are what come back.
+const ITEM_ENTRY: u32 = 7973;
+
+pub(crate) struct OpenItem;
+
+impl Probe for OpenItem {
+    fn stage(&mut self, cx: &mut Ctx) -> Result<()> {
+        cx.session.send_chat(&format!(".additem {ITEM_ENTRY}"))?;
+        println!("sent GM: .additem {ITEM_ENTRY}");
+        Ok(())
+    }
+
+    fn verify(&mut self, cx: &mut Ctx) -> Result<()> {
+        let world = &mut *cx.world;
+        let session = &mut *cx.session;
+
+        // 1) Find the clam in the backpack, and remember WHICH slot — the wire addresses the item
+        // by position (bag 255 + the absolute player-array slot), never by guid.
+        let sf = world
+            .self_fields
+            .as_ref()
+            .context("no self descriptor — can't walk the backpack")?;
+        let (slot0, item_guid) = (0..16)
+            .filter_map(|i| sf.player_pack_slot(i).filter(|g| *g != 0).map(|g| (i, g)))
+            .find(|(_, g)| world.item_entries.get(g).copied() == Some(ITEM_ENTRY))
+            .with_context(|| {
+                format!(
+                    "--open-item: item {ITEM_ENTRY} isn't in the backpack — did `.additem` land? \
+                     (needs a GM account; try a longer --seconds)"
+                )
+            })?;
+        let wire_slot = 23 + slot0;
+        println!(
+            "\nopenable item {ITEM_ENTRY}: guid {item_guid:#x} in pack slot {} → \
+             CMSG_OPEN_ITEM bag 255 slot {wire_slot}",
+            slot0 + 1
+        );
+
+        // 2) The fork's own packet. A loot response on the ITEM's guid is the proof: nothing else
+        // in the protocol opens a loot window over a bag position.
+        session.open_item(255, wire_slot)?;
+        let drain_until = Instant::now() + Duration::from_secs(5);
+        let mut verdict: Option<Result<String, String>> = None;
+        while Instant::now() < drain_until && verdict.is_none() {
+            let Ok(msg) = session.recv() else { continue };
+            for ev in decode(msg) {
+                match ev {
+                    SessionEvent::LootResponse {
+                        guid,
+                        loot_type,
+                        gold,
+                        items,
+                    } if guid == item_guid => {
+                        verdict = Some(Ok(format!(
+                            "SMSG_LOOT_RESPONSE on the ITEM guid — type {loot_type}, {} copper, \
+                             {} row(s): {}",
+                            gold,
+                            items.len(),
+                            items
+                                .iter()
+                                .map(|i| format!("{}x{}", i.count, i.item_id))
+                                .collect::<Vec<_>>()
+                                .join(", ")
+                        )));
+                    }
+                    SessionEvent::LootError { guid, error } if guid == item_guid => {
+                        verdict = Some(Err(format!("server refused with loot error {error}")));
+                    }
+                    SessionEvent::InventoryFailure { reason, .. } => {
+                        verdict = Some(Err(format!(
+                            "server refused with EQUIP_ERR {reason} (locked? dead? flying?)"
+                        )));
+                    }
+                    _ => {}
+                }
+            }
+        }
+        let v = verdict
+            .context("--open-item: no reaction to CMSG_OPEN_ITEM within 5s")?
+            .map_err(|e| anyhow::anyhow!("--open-item: {e}"))?;
+        println!("✅ open item: {v}");
+
+        // 3) Close the window again — an abandoned loot session would follow the character.
+        session.loot_release(item_guid)?;
+        let until = Instant::now() + Duration::from_secs(2);
+        while Instant::now() < until {
+            let Ok(msg) = session.recv() else { continue };
+            for ev in decode(msg) {
+                if let SessionEvent::LootReleaseResponse { guid } = ev {
+                    if guid == item_guid {
+                        println!("✅ release: SMSG_LOOT_RELEASE_RESPONSE on {guid:#x}");
+                    }
+                }
+            }
+        }
+
+        // Leave the probe character as found (the clam survives an un-emptied open).
+        session.send_chat(&format!(".additem {ITEM_ENTRY} -1"))?;
+        println!("cleanup: .additem {ITEM_ENTRY} -1");
+        Ok(())
+    }
+}
@@ -191,8 +191,45 @@ impl ItemInfo {
    pub fn placeable_on_action_bar(&self) -> bool {
        self.use_spell.is_some() || self.inventory_type != 0
    }
+
+    /// Does a right-click **open** this concrete instance rather than use it? `instance_flags` is
+    /// the item object's `ITEM_FIELD_FLAGS` (0 for a template-only view, which can never be
+    /// openable — you can't right-click a hyperlink).
+    ///
+    /// The one predicate behind both halves of the affordance: the tooltip's green
+    /// `<Right Click to Open>` line and the `CMSG_OPEN_ITEM` fork. It is the reference's own
+    /// tooltip condition (wow-re `tooltip-content-law.md` §1-OPENABLE, byte-read at
+    /// `0x52e2f8`–`0x52e321`): the template LOOTABLE bit [`ITEM_FLAG_LOOTABLE`] behind its lock
+    /// sub-gate — a [`ItemInfo::lock_id`] item only opens once the instance carries
+    /// [`ITEM_DYNFLAG_UNLOCKED`] — **or** a wrapped gift ([`ITEM_FLAG_WRAPPER`] on the template
+    /// plus [`ITEM_DYNFLAG_WRAPPED`] on the instance). vmangos's `HandleOpenItemOpcode` gates on
+    /// exactly the same pair, so the two ends agree: what the tooltip promises is what the server
+    /// will honour, and a still-locked junkbox neither advertises the line nor sends the packet.
+    pub fn openable(&self, instance_flags: u32) -> bool {
+        let lootable = self.flags & ITEM_FLAG_LOOTABLE != 0
+            && (self.lock_id == 0 || instance_flags & ITEM_DYNFLAG_UNLOCKED != 0);
+        let gift =
+            self.flags & ITEM_FLAG_WRAPPER != 0 && instance_flags & ITEM_DYNFLAG_WRAPPED != 0;
+        lootable || gift
+    }
 }

+/// `ITEM_FLAG_LOOTABLE` — the **template** bit that makes an item right-clickable into a loot
+/// window (a clam, a lockbox, a Gnomish Mind Control Cap's box). vmangos
+/// `ItemPrototype.h:66`, whose own comment names the client behaviour this drives: "It or lockid
+/// set enable for client show 'Right click to open'".
+pub const ITEM_FLAG_LOOTABLE: u32 = 0x0000_0004;
+/// `ITEM_FLAG_WRAPPER` — the template bit for gift wrapping (vmangos `ItemPrototype.h:73`); paired
+/// with [`ITEM_DYNFLAG_WRAPPED`] on the instance it makes the wrapped present openable.
+pub const ITEM_FLAG_WRAPPER: u32 = 0x0000_0200;
+/// `ITEM_DYNFLAG_UNLOCKED` — the **instance** (`ITEM_FIELD_FLAGS`) bit a lockbox gains once it has
+/// been picked/keyed open; until then a `LockID` item refuses to open (vmangos
+/// `HandleOpenItemOpcode`'s `EQUIP_ERR_ITEM_LOCKED`).
+pub const ITEM_DYNFLAG_UNLOCKED: u32 = 0x0000_0004;
+/// `ITEM_DYNFLAG_WRAPPED` — the instance bit on a gift-wrapped item; opening it unwraps to the
+/// present (vmangos swaps the entry from `character_gifts` rather than sending loot).
+pub const ITEM_DYNFLAG_WRAPPED: u32 = 0x0000_0008;
+
 /// One `Damage` block ([`ItemInfo::damages`] — block 0 is also mirrored into
 /// [`ItemInfo::dmg_min`]/[`ItemInfo::dmg_max`]/[`ItemInfo::dmg_type`]).
 #[derive(Debug, Clone, Copy, PartialEq)]
@@ -508,6 +545,18 @@ pub fn use_item(bag_index: u8, slot: u8, spell_slot: u8, go_target: Option<u64>)
    body
 }

+/// Body of `CMSG_OPEN_ITEM` (VERIFIED vmangos `OpenItem::ReadFromWorldPacket`,
+/// `Server/Packets/Spell.cpp` + `.h:36-45`; opcode 172 `Opcodes_1_12_1.h:175`): `bagIndex`, `slot`
+/// — the same two bytes and the same bag addressing as [`use_item`], and nothing else (no spell
+/// ordinal, no targets block: opening is not a cast).
+///
+/// The right-click fork for an [`ItemInfo::openable`] item. The server answers on the **item's own
+/// guid**: `SendLoot(item, LOOT_CORPSE)` for a lootable, or — for a wrapped gift — an entry swap
+/// and no window at all.
+pub fn open_item(bag_index: u8, slot: u8) -> Vec<u8> {
+    vec![bag_index, slot]
+}
+
 /// Body of `CMSG_AUTOEQUIP_ITEM` (VERIFIED vmangos `AutoEquipItem::ReadFromWorldPacket`,
 /// `Server/Packets/Item.cpp:17-21` + `.h:31-39`; opcode 266 `Opcodes_1_12_1.h:269`): source
 /// `srcbag`/`srcslot` (both `uint8`), the same bag addressing as [`use_item`]. The real client
@@ -105,9 +105,10 @@ pub use group::{
    PartyMemberStatsInfo, RaidTargetUpdate, ReadyCheck, GROUP_MEMBER_ASSISTANT,
 };
 pub use items::{
-    auto_equip_item, auto_store_bag_item, destroy_item, item_query, set_ammo, split_item,
-    swap_inv_item, swap_item, use_item, ItemDamage, ItemInfo, ItemSpellEntry, ItemUseSpell,
-    BAG_PLAYER_INVENTORY, SLOT_BAG_FIRST, SLOT_PACK_FIRST,
+    auto_equip_item, auto_store_bag_item, destroy_item, item_query, open_item, set_ammo,
+    split_item, swap_inv_item, swap_item, use_item, ItemDamage, ItemInfo, ItemSpellEntry,
+    ItemUseSpell, BAG_PLAYER_INVENTORY, ITEM_DYNFLAG_UNLOCKED, ITEM_DYNFLAG_WRAPPED,
+    ITEM_FLAG_LOOTABLE, ITEM_FLAG_WRAPPER, SLOT_BAG_FIRST, SLOT_PACK_FIRST,
 };
 pub use loot::{
    autostore_loot_item, loot, loot_error, loot_money, loot_release, loot_roll, loot_type,
@@ -218,6 +218,12 @@ pub const CMSG_GAMEOBJECT_QUERY: u16 = 0x005E; // 94
 pub const CMSG_QUEST_QUERY: u16 = 0x005C; // 92
 pub const CMSG_ITEM_QUERY_SINGLE: u16 = 0x0056; // 86
 pub const CMSG_USE_ITEM: u16 = 0x00AB; // 171
+/// VERIFIED vmangos `Opcodes_1_12_1.h:175`: 172. Body in [`super::items::open_item`] — the bag
+/// position of an *openable* item (a clam, an unlocked lockbox, a wrapped gift). The right-click
+/// fork `CMSG_USE_ITEM` never takes: the server answers `SMSG_LOOT_RESPONSE` on the **item's own
+/// guid** (`HandleOpenItemOpcode` → `SendLoot(item, LOOT_CORPSE)`), so the loot window opens over
+/// a thing in your bag rather than a corpse in the world.
+pub const CMSG_OPEN_ITEM: u16 = 0x00AC; // 172
 /// VERIFIED vmangos `Opcodes_1_12_1.h`: 177. Body in [`super::gameobj_use`] — a full guid, the
 /// GameObject to use (decision 0236). Not interchangeable with `CMSG_LOOT`: the server rejects a
 /// GameObject guid on `CMSG_LOOT`, so a chest opens its loot through this opcode.
@@ -360,6 +360,15 @@ impl WorldSession {
        )
    }

+    /// Open an item by bag position (`CMSG_OPEN_ITEM`) — the unsplit twin of
+    /// [`WorldWriter::open_item`], for the probe's live open-item verification.
+    pub fn open_item(&mut self, bag_index: u8, slot: u8) -> Result<()> {
+        self.send(
+            opcode::CMSG_OPEN_ITEM,
+            &messages::open_item(bag_index, slot),
+        )
+    }
+
    /// Equip a bag item (`CMSG_AUTOEQUIP_ITEM`) — the unsplit twin of
    /// [`WorldWriter::auto_equip_item`], for the probe's live equip verification.
    pub fn auto_equip_item(&mut self, bag_index: u8, slot: u8) -> Result<()> {
@@ -43,6 +43,18 @@ impl WorldWriter {
        )
    }

+    /// Open an item by bag position (`CMSG_OPEN_ITEM`, layout in [`messages::open_item`]) — crack
+    /// the clam, empty the picked lockbox, unwrap the gift. The right-click fork for an
+    /// [`crate::ItemInfo::openable`] item; the server answers with `SMSG_LOOT_RESPONSE` on the
+    /// item's **own** guid (so the loot window opens over a thing in the bag), or an equip error
+    /// on a refusal (still locked, dead, flying).
+    pub fn open_item(&mut self, bag_index: u8, slot: u8) -> Result<()> {
+        self.send(
+            opcode::CMSG_OPEN_ITEM,
+            &messages::open_item(bag_index, slot),
+        )
+    }
+
    /// Equip a bag item (`CMSG_AUTOEQUIP_ITEM`, layout in [`messages::auto_equip_item`]) — the
    /// server picks the destination slot. Success arrives as inventory-slot values deltas (and the
    /// visible-item change everyone renders); refusal as `SMSG_INVENTORY_CHANGE_FAILURE`.
@@ -42,6 +42,15 @@ fn item_query_wire() {
    // CMSG_AUTOEQUIP_ITEM (vmangos AutoEquipItem::ReadFromWorldPacket): bagIndex, slot.
    assert_eq!(messages::auto_equip_item(255, 25), hx("ff19"));

+    // CMSG_OPEN_ITEM (vmangos OpenItem::ReadFromWorldPacket, Server/Packets/Spell.cpp:19-23):
+    // bagIndex, slot — and NOTHING else (no spell ordinal, no targets block; opening is not a
+    // cast). The same two bytes as auto-equip under a different opcode: this is the fork a
+    // clam/lockbox/gift takes instead of USE_ITEM, and the only one the server answers with
+    // SendLoot on the item's own guid. Backpack slot 3 = bag 255 + player-array slot 25; an item
+    // inside an equipped bag addresses that bag's own player-array slot (19..22) + inner slot.
+    assert_eq!(messages::open_item(255, 25), hx("ff19"));
+    assert_eq!(messages::open_item(19, 0), hx("1300"));
+
    // SMSG_INVENTORY_CHANGE_FAILURE (vmangos InventoryChangeFailure::AppendBodyTo): the level
    // refusal (reason 1) carries a u32 required level before the guids; others don't.
    let level_fail = messages::parse_server(
@@ -306,6 +306,53 @@ impl ZKey {
    pub const fn raw(self) -> u64 {
        self.0
    }
+
+    /// Unpack this key back into its fields.
+    #[inline]
+    pub const fn parts(self) -> ZParts {
+        unpack(self.0)
+    }
+}
+
+/// The fields of a packed [`ZKey`], in significance order — the readable face of an opaque `u64`.
+///
+/// A draw-order key is a single number by design (one sort produces the whole render list), which
+/// makes every question *about* an order — "why is this over that?", "which pairs would a re-rank
+/// invert?" — unanswerable at a glance. [`unpack`] is that answer, and it is why
+/// [`crate::script::ExtractedQuad::z`] can stay a bare `u64`.
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub struct ZParts {
+    /// The frame strata bucket id (0..=9, [`Strata::index`]).
+    pub strata: u8,
+    /// The frame level within the strata bucket.
+    pub level: u16,
+    /// The draw layer (0..=4, [`DrawLayer::index`]) — bucket-wide, above the frame (decision 0884).
+    pub layer: u8,
+    /// `true` for a font string: all textures of a `(strata, level, layer)` precede all its text.
+    pub is_fontstring: bool,
+    /// The owning frame's link-stamp — its position in the client's intrusive bucket list.
+    pub insertion: u32,
+    /// `false` for a frame's own slot, `true` for one of its regions.
+    pub is_region: bool,
+    /// The Era `textureSubLevel` knob; inert for 1.12 content (always 0 there).
+    pub sub_level: i8,
+    /// Declaration order within the owning frame's layer.
+    pub decl: u16,
+}
+
+/// Unpack a raw [`ZKey`] (e.g. [`crate::script::ExtractedQuad::z`]) into its fields.
+#[inline]
+pub const fn unpack(raw: u64) -> ZParts {
+    ZParts {
+        strata: ((raw >> STRATUM_SHIFT) & 0xf) as u8,
+        level: ((raw >> LEVEL_SHIFT) & 0xffff) as u16,
+        layer: ((raw >> LAYER_SHIFT) & 0x7) as u8,
+        is_fontstring: (raw >> FONTSTRING_SHIFT) & 1 == 1,
+        insertion: ((raw >> INSERTION_SHIFT) & ((1 << INSERTION_BITS) - 1)) as u32,
+        is_region: (raw >> IS_REGION_SHIFT) & 1 == 1,
+        sub_level: (((raw >> SUBLEVEL_SHIFT) & 0xff) as i16 - 128) as i8,
+        decl: ((raw >> DECL_SHIFT) & ((1 << DECL_BITS) - 1)) as u16,
+    }
 }

 // ─────────────────────────────────────────────────────────────────────────────────────────────
@@ -329,9 +376,9 @@ pub enum ZTarget {
 /// (`propagation.md`) — a child of a hidden frame already carries `effective_visible == false` and is
 /// skipped here. Each visible frame emits its own [`ZTarget::Frame`] entry followed by one
 /// [`ZTarget::Region`] per owned region; the returned vec is sorted ascending by [`ZKey`], which *is*
-/// the total draw order (strata → **layer** → frame link-stamp → is-region → sub-level →
-/// texture<fontstring → decl; the layer outranks the frame — see the `ZKey` bit-layout note and
-/// decision 0884). Ties are impossible: distinct frames differ in link-stamp, a frame and its
+/// the total draw order (strata → frame level → **draw layer** → texture<fontstring → frame
+/// link-stamp → is-region → sub-level → decl; the layer outranks the frame — see the `ZKey`
+/// bit-layout note and decision 0884). Ties are impossible: distinct frames differ in link-stamp, a frame and its
 /// regions differ in the is-region bit, and a frame's regions differ in the remaining fields.
 ///
 /// Ordering only: this emits an entry for *every* region of a visible frame. Region-level
@@ -425,8 +425,9 @@ fn required_level_one_is_hidden() {
 /// (white) when the instance carries letter text, the green-escaped "<Made by %s>" otherwise —
 /// then READABLE off the instance text id (its template's PageText is 0 — the
 /// director-reported gap). An unresolved creator (name query in flight) emits no line. LOCKED
-/// yields to the instance UNLOCKED bit — and ITEM_OPENABLE never emits from `SetBagItem` (a
-/// p6=1 source: the `0x52e2e8` bypass), so the unlocked chest shows its bare name.
+/// yields to the instance UNLOCKED bit — which is also what un-gates ITEM_OPENABLE: a bag hover
+/// on an openable instance DOES show the green `<Right Click to Open>` (director-observed on a
+/// clam; see `ItemInstance::openable_source` for why this overrides §1-OPENABLE's p6 reading).
 #[test]
 fn instance_tail_creator_and_readable() {
    let mut s = script();
@@ -458,16 +459,26 @@ fn instance_tail_creator_and_readable() {
        flags,
        ..Default::default()
    };
+    s.set_item_template(
+        7973,
+        ItemTemplateView {
+            name: "Small Barnacled Clam".into(),
+            quality: 1,
+            flags: 0x4,
+            ..Default::default()
+        },
+    );
    let mut slots = HashMap::new();
    slots.insert(1, slot(889, true, Some("One"), 0)); // the mail letter
    slots.insert(2, slot(889, true, None, 0)); // creator name still in flight
    slots.insert(3, slot(2589, false, Some("Geoffrey"), 0)); // crafted, still locked
-    slots.insert(4, slot(2589, false, None, 0x4)); // unlocked → LOCKED gone, and NO open line
+    slots.insert(4, slot(2589, false, None, 0x4)); // unlocked → LOCKED gone, open line on
+    slots.insert(5, slot(7973, false, None, 0)); // the clam: lockless, openable outright
    s.set_container(
        0,
        Some(ContainerState {
            name: Some("Backpack".into()),
-            num_slots: 4,
+            num_slots: 5,
            slots,
        }),
    );
@@ -502,11 +513,18 @@ fn instance_tail_creator_and_readable() {
        vec!["Heavy Chest", "Locked", "|cff00ff00<Made by Geoffrey>|r"],
        "CREATED_BY carries the string's own green escape; locked chest hides OPENABLE"
    );
-    // The instance UNLOCKED bit retires the LOCKED line — and OPENABLE still never emits
-    // (SetBagItem is p6=1; only send-mail/auction/buyback sources reach the emitter in 5875).
+    // The instance UNLOCKED bit retires the LOCKED line AND satisfies the openable lock sub-gate.
    s.run(r#"TT:SetOwner(getglobal("SlotL"), "ANCHOR_RIGHT"); TT:SetBagItem(0, 4)"#)
        .unwrap();
+    let lines = lines_of(&mut s);
+    let texts: Vec<&str> = lines.iter().map(|(t, _)| t.as_str()).collect();
+    assert_eq!(texts, vec!["Heavy Chest", "<Right Click to Open>"]);
+    assert_eq!(lines[1].1, [0.0, 1.0, 0.0, 1.0], "OPENABLE is green");
+    // The director's case: a lockless LOOTABLE template (a clam) is openable outright — name +
+    // the green line, nothing between them.
+    s.run(r#"TT:SetOwner(getglobal("SlotL"), "ANCHOR_RIGHT"); TT:SetBagItem(0, 5)"#)
+        .unwrap();
    let texts: Vec<String> = lines_of(&mut s).into_iter().map(|(t, _)| t).collect();
-    assert_eq!(texts, vec!["Heavy Chest"]);
+    assert_eq!(texts, vec!["Small Barnacled Clam", "<Right Click to Open>"]);
    assert!(s.take_errors().is_empty());
 }
@@ -309,7 +309,10 @@ pub(super) fn install_methods(lua: &Lua, m: &Table) -> mlua::Result<()> {
                                creator: s.creator.clone(),
                                has_text: false,
                                flags: s.flags,
-                                // p6=1 in the ref — SetInventoryItem can never emit OPENABLE.
+                                // Unobservable either way: nothing openable is equippable, so the
+                                // doll hover has no case to show. Left `false` (wow-re
+                                // §1-OPENABLE's p6 reading) rather than flipped on the bag's
+                                // evidence — see `ItemInstance::openable_source`.
                                openable_source: false,
                            },
                        )
@@ -375,8 +378,11 @@ pub(super) fn install_methods(lua: &Lua, m: &Table) -> mlua::Result<()> {
                            creator: s.creator.clone(),
                            has_text: s.readable,
                            flags: s.flags,
-                            // p6=1 in the ref — SetBagItem can never emit OPENABLE (§1-OPENABLE).
-                            openable_source: false,
+                            // The bag hover DOES emit `<Right Click to Open>` — director-observed
+                            // (a clam in a bag), against wow-re §1-OPENABLE's p6 reading. See
+                            // `ItemInstance::openable_source`: that gate is under re-derivation,
+                            // and the observation is the ground truth we build on.
+                            openable_source: true,
                        },
                    ),
                    None => return Ok(MultiValue::from_vec(vec![Value::Nil])),
@@ -29,11 +29,24 @@ pub(super) struct ItemInstance {
    /// Instance `ITEM_FIELD_FLAGS` — the openable lock sub-gate reads UNLOCKED `0x4`
    /// (`0x52e30c`) and the wrapped-gift arm WRAPPED `0x8` (`0x52e31d`).
    pub flags: u32,
-    /// The ref's `p6=0` source class (wow-re §1-OPENABLE, byte-census 2026-07-20): ITEM_OPENABLE
-    /// is reachable ONLY from object-present p6=0 sources (SetSendMailItem, SetAuctionSellItem,
-    /// SetBuybackItem) — `SetBagItem`/`SetInventoryItem` pass p6=1 and a `[this+0x440]` gate
-    /// (`0x52e2e8`) bypasses the openable tree entirely. Every current benilla source is p6=1
-    /// or template-only, so this stays `false` until those tooltips carry real instances.
+    /// May this source emit the ITEM_OPENABLE line at all?
+    ///
+    /// wow-re §1-OPENABLE (byte-census 2026-07-20) reads a `[this+0x440]` gate at `0x52e2e8` as
+    /// bypassing the whole openable tree for every `p6=1` binding, and concludes that
+    /// `SetBagItem` — the binding your own bags hover through — can NEVER emit it. **That
+    /// conclusion is contradicted by ground truth:** the director's screenshot shows a Small
+    /// Barnacled Clam hovered IN A BAG with the green `<Right Click to Open>` line, and vmangos's
+    /// own `ITEM_FLAG_LOOTABLE` comment names that exact client behaviour ("It or lockid set
+    /// enable for client show 'Right click to open'"). A line that could only ever appear on the
+    /// send-mail / auction-sell / buyback tooltips would also be behaviourally absurd. The note is
+    /// additionally self-inconsistent there (§1-CREATOR puts the GIFTCREATOR read at `0x52e32a`,
+    /// inside the region §1-OPENABLE calls the openable emit + jump), so the transcription — not
+    /// the observation — is the suspect. A wow-re §5 re-derivation of `[0x52e2e0, 0x52e376)` is in
+    /// flight; until it lands, the bag hover follows the director's eyes.
+    ///
+    /// The flag is kept rather than deleted because the *template-vs-instance* distinction it also
+    /// carries is settled (gate 1, `0x52e2e0`: no object ⇒ no line — you can't right-click a
+    /// hyperlink), and because the p6 split may still be real for the sources we cannot observe.
    pub openable_source: bool,
 }

@@ -67,6 +67,13 @@
            frame.portrait   = getglobal(name .. "Portrait")
            frame.healthBar  = getglobal(name .. "HealthBar")
            frame.manaBar    = getglobal(name .. "ManaBar")
+            -- The recessed-bar look, the reference's own mechanism (ref-TargetFrame.lua l.32-34):
+            -- push the bars one frame level BELOW the art's TextureFrame. Frame level is the only
+            -- key term above the draw layer (decision 0884), so this is the only thing that can hold
+            -- the BACKGROUND party art over the ARTWORK bar fills. See the LAYERING note below.
+            local textureFrameLevel = getglobal(tf):GetFrameLevel()
+            frame.healthBar:SetFrameLevel(textureFrameLevel - 1)
+            frame.manaBar:SetFrameLevel(textureFrameLevel - 1)
            frame.nameText   = getglobal(tf .. "Name")
            frame.leaderIcon = getglobal(tf .. "LeaderIcon")
            frame.masterIcon = getglobal(tf .. "MasterIcon")
@@ -257,12 +264,18 @@
    <!-- The shared member-frame template — ref PartyFrameTemplates.xml, 128x53 over the 128x64
         UI-PartyFrame art at (0,-2) (the art hangs 13 px below the frame rect, as authored).

-         LAYERING (the ref's law, in our z-model's idiom): the ref nests the frame art two anonymous
-         child frames deep so it draws OVER the status bars — the bars show through the art's
-         semi-transparent tube windows, inset, their ends clipped by the rails. Ours is the
-         UnitFrames.xml idiom: a $parentTextureFrame child declared AFTER the bars. The name and the
-         leader/master/disconnect icons ride in it too (the ref keeps them in that nested frame;
-         left below the art, the ring would swallow the leader crown). -->
+         LAYERING (the ref's law): the ref nests the frame art two anonymous child frames deep so it
+         draws OVER the status bars — the bars show through the art's semi-transparent tube windows,
+         inset, their ends clipped by the rails. Ours is the UnitFrames.xml idiom, which reaches the
+         same place by the ref's OTHER route (its TargetFrame's): a $parentTextureFrame child, with
+         OnLoad pushing the bars one frame level BELOW it. Both spend a frame LEVEL, which is the
+         point — level is the one key term above the draw layer (decision 0884). Until then this file
+         instead leaned on declaring the TextureFrame after the bars, and the old draw key's
+         insertion-order tie-break; 0884 pinned the layer as bucket-wide and above the frame, which
+         put the bars' ARTWORK fills straight over the BACKGROUND art. Do not go back to declaration
+         order. The name and the leader/master/disconnect icons ride in the TextureFrame too (the ref
+         keeps them in that nested frame; left below the art, the ring would swallow the leader
+         crown). -->
    <!-- frameStrata LOW is the ref's own (`PartyFrameTemplates.xml` l.194) and it is load-bearing,
         not decoration. The panel windows (loot, character, merchant, quest log, …) are MEDIUM, and
         the party frames share the top-left corner with the left-slot panel. Left in the default
@@ -12,12 +12,20 @@
        LAYERING (matches the ref's structure). The draw order must be trough < bars < ring art < text:
        the frame's metal well-edges overlap the bar edges so the bars read RECESSED, not pasted on top.
        The ref gets this with a child `TargetFrameTextureFrame` (ring art) and `SetFrameLevel(bar,
-        textureFrame-1)` to push the bars beneath it. Our z-model tie-breaks a `(strata, level)` bucket by
-        frame *insertion* order (crates/benilla-ui/src/order.rs), so we get the same result structurally
-        and without SetFrameLevel: the parent holds trough + portrait (+ target's name plate) in
-        BACKGROUND; the `$parentHealthBar`/`$parentPowerBar` StatusBars come next; then
+        textureFrame-1)` to push the bars beneath it, and we do the same, in BenillaUnitFrame_OnLoad:
+        the parent holds trough + portrait (+ target's name plate) in BACKGROUND; the
+        `$parentHealthBar`/`$parentPowerBar` StatusBars sit at the PARENT's frame level; and
        `$parentTextureFrame` (232×100, filling the parent) holds the ring art (BACKGROUND) + name/level/dead text
-        (OVERLAY), declared LAST so it draws over the bar edges and the portrait. The ring's art is a THIN
+        (OVERLAY) one level ABOVE them, so it draws over the bar edges and the portrait.
+
+        That SetFrameLevel is load-bearing, and this file used to do without it. Until decision 0884
+        the bars were held under the ring art by DECLARING `$parentTextureFrame` after them, leaning on
+        the old draw key's insertion-order tie-break between siblings at one `(strata, level)`. 0884
+        pinned the real law at the binary: the draw LAYER is bucket-wide and outranks the frame, so
+        siblings interleave by layer and the bars' ARTWORK fills painted straight over the TextureFrame's
+        BACKGROUND ring art (the frames read as pasted-on slabs — director report, 2026-08-02). Frame
+        level is the one term above the layer, which is exactly why the reference spends a level here.
+        Do not "simplify" this back into declaration order. The ring's art is a THIN
        band with transparent corners (not an opaque surround), so the square is hidden by the portrait
        itself — set via SetPortraitToTexture, masked to its inscribed circle (the client's portraits are
        always round: a model render or a circular-cropped icon), so no corner reaches the ring's gap. The
@@ -129,6 +137,13 @@
            -- The name/level/dead text live inside $parentTextureFrame (so they draw over the ring art
            -- that draws over the bars), so their global names carry the TextureFrame prefix.
            local tf = name .. "TextureFrame"
+            -- The recessed-bar look, the reference's own mechanism (ref-TargetFrame.lua l.32-34):
+            -- push the bars one frame level BELOW the ring art's TextureFrame. Frame level is the
+            -- only key term above the draw layer (decision 0884), so this is the only thing that can
+            -- hold a BACKGROUND ring texture over an ARTWORK bar fill. See the LAYERING note up top.
+            local textureFrameLevel = getglobal(tf):GetFrameLevel()
+            frame.healthBar:SetFrameLevel(textureFrameLevel - 1)
+            frame.powerBar:SetFrameLevel(textureFrameLevel - 1)
            frame.nameText  = getglobal(tf .. "NameText")
            frame.levelText = getglobal(tf .. "LevelText")
            frame.deadText  = getglobal(tf .. "DeadText")
@@ -752,8 +767,11 @@
                    </Anchors>
                    <Color r="0" g="0" b="0" a="0.5"/>
                </Texture>
-                <!-- Portrait 64×64 TOPLEFT (42,-12) (ref l.35-46, moved ARTWORK→BACKGROUND so the ring
-                     art draws over it). Texture set in OnLoad via SetPortraitToTexture (circle-masked). -->
+                <!-- Portrait 64×64 TOPLEFT (42,-12) (ref l.35-46, which puts it in ARTWORK; ours sits
+                     in BACKGROUND, a leftover of the pre-0884 layering. The ring art now clears it by
+                     frame LEVEL, and nothing else in this layer overlaps the portrait's square, so the
+                     two are indistinguishable here — left alone rather than churned inside a bug fix).
+                     Texture set in OnLoad via SetPortraitToTexture (circle-masked). -->
                <Texture name="$parentPortrait">
                    <Size><AbsDimension x="64" y="64"/></Size>
                    <Anchors>
@@ -761,8 +779,9 @@
                    </Anchors>
                </Texture>
            </Layer>
-            <!-- The ring art + name/level text live in $parentTextureFrame below (declared AFTER the
-                 bars) so they draw OVER the bar edges — the recessed-bar look. See the LAYERING note. -->
+            <!-- The ring art + name/level text live in $parentTextureFrame below, one frame level
+                 ABOVE the bars, so they draw OVER the bar edges — the recessed-bar look. OnLoad puts
+                 them there (SetFrameLevel); see the LAYERING note at the top of this file. -->
        </Layers>
        <Frames>
            <!-- Health bar 119×12 TOPLEFT (106,-41), UI-StatusBar, green (ref l.288-315). -->
@@ -784,8 +803,9 @@
                <BarColor r="0.0" g="0.0" b="1.0"/>
            </StatusBar>
            <!-- The ring art + text, over the bars (ref TargetFrameTextureFrame). 232×100 fills the
-                 parent; declared after the StatusBars so it draws over their edges — our z-model's
-                 insertion-order equivalent of the ref's SetFrameLevel(bar, textureFrame-1). -->
+                 parent; it draws over the bar edges because OnLoad pushes the bars one frame level
+                 BELOW it (the ref's SetFrameLevel(bar, textureFrame-1)) — NOT because of where it
+                 is declared. See the LAYERING note at the top of this file. -->
            <Frame name="$parentTextureFrame">
                <Size><AbsDimension x="232" y="100"/></Size>
                <Anchors>
@@ -965,9 +985,10 @@
                <BarTexture file="Interface\TargetingFrame\UI-StatusBar"/>
                <BarColor r="0.0" g="0.0" b="1.0"/>
            </StatusBar>
-            <!-- The ring art + text, over the bars (ref TargetFrameTextureFrame); declared after the
-                 StatusBars so it draws over their edges (our z-model's insertion-order equivalent of
-                 the ref's SetFrameLevel(bar, textureFrame-1)). -->
+            <!-- The ring art + text, over the bars (ref TargetFrameTextureFrame); it draws over the
+                 bar edges because OnLoad pushes the bars one frame level BELOW it (the ref's
+                 SetFrameLevel(bar, textureFrame-1)) — NOT because of where it is declared. See the
+                 LAYERING note at the top of this file. -->
            <Frame name="$parentTextureFrame">
                <Size><AbsDimension x="232" y="100"/></Size>
                <Anchors>
@@ -85,13 +85,14 @@
 //!   [`crate::model_fade::ParentModel`] link up to the unit — the recursion itself, and the same walk
 //!   `ModelAlphas` already does for alpha.
 //!
-//! **Proc 11 ships too** (decision 0889): the **freeze**. Its param is a playback *rate* written onto
-//! the unit's own animation clocks (`0x60db7e` → `SetBoneAnimSpeed 0x712910` on the mount's bone 0,
-//! the body's key-bone 4 — the upper-body split — and the body's bone 0), old rates saved in the node
-//! and written back when it expires (`0x6203e0`). Rate 0 is Ice Block: the pose holds exactly where
-//! it was, *including* a cast one-shot that had just been armed, which is why the caster never gets
-//! to raise their hand. [`apply_aura_anim_rate`] is that leg, and the reference's save/restore is
-//! [`AnimRateFreeze`].
+//! **Proc 11 ships too** (decisions 0889 + 0891): the **freeze**. Its param is a playback *rate*
+//! written onto the unit's own animation clocks (`0x60db7e` → `SetBoneAnimSpeed 0x712910` on the
+//! mount's bone 0, the body's key-bone 4 — the upper-body split — and the body's bone 0), old rates
+//! saved in the node and written back when it expires (`0x6203e0`). Rate 0 is Ice Block: the pose
+//! holds exactly where it was, *including* a cast one-shot that had just been armed, which is why the
+//! caster never gets to raise their hand. [`apply_aura_anim_rate`] is that leg — and it expresses
+//! rate 0 as a **pause**, not as a speed, because the reference's rate belongs to a bone while a speed
+//! here belongs to a clip that other code copies from ([`AnimRateFreeze`] carries the story).
 //!
 //! Types 2, 7 and 13 also appear on state kits (Berserk/Bloodlust's 2, the Sap/Feign-Death 7, the
 //! "Glowy (Red)" 13) and have **no verified mechanism** — they fall through [`node_for`]'s single `_`
@@ -319,33 +320,50 @@ fn tint_probe() -> Option<[u8; 3]> {
    })
 }

-/// On a rig whose clocks a proc-11 aura is holding: the playback speeds we took over, so the release
-/// can hand them back — the reference's own `+0x60`/`+0x64`/`+0x68` (`0x6201d0` saves the three bone
-/// rates before overwriting them; `0x6203e0` writes them back when the node expires with the aura).
+/// Marker: this rig's clocks are being held by a proc-11 aura, so the release knows to let them go.
 ///
-/// Keyed by animation node because that is what a speed belongs to here: the reference saves per
-/// *bone* because a rate is a bone-clock property there, and ours is per armed clip.
+/// It carries **nothing**, and that is the fix for the first shipped version's bug. That one stored
+/// the speed of every clip it took over and wrote the values back on release — the reference's own
+/// `+0x60`/`+0x64`/`+0x68` save/restore, transliterated one level too low. It cannot work here,
+/// because in the reference a rate belongs to a *bone* (one value, saved once, and clips armed later
+/// never carry one) while ours would have to belong to a *clip*, and clips get armed under the freeze
+/// carrying values copied from other, already-frozen clips. `transplant_up` is exactly that: moving a
+/// one-shot to the torso overlay copies the source clip's speed, which under the freeze was `0`, so
+/// the overlay was saved at `0`, restored to `0`, and never advanced again — a cast pose welded to the
+/// upper body for the rest of the session, surviving every later cast (it never *finished*, so the
+/// overlay never released). Pausing owns no value and so cannot poison one.
 #[derive(Component, Default, Debug)]
-pub(crate) struct AnimRateFreeze {
-    saved: Vec<(AnimationNodeIndex, f32)>,
-}
+pub(crate) struct AnimRateFreeze;

-/// Hold every unit whose head proc-11 node says so at that node's rate — its own rig and its mount's
-/// (`[unit+0xd8]` and `[unit+0xdc]`, the two models `0x6201d0` writes) — and hand the speeds back
-/// when the aura goes.
+/// Hold the clocks of every unit whose head proc-11 node says so — its own rig and its mount's
+/// (`[unit+0xd8]` and `[unit+0xdc]`, the two models `0x6201d0` writes) — and let them go when the aura
+/// does.
+///
+/// **A rate of 0 is expressed as a PAUSE, not as `set_speed(0.0)`.** They are the same picture on
+/// screen (Bevy evaluates a paused animation at its frozen seek, exactly as `0x712910`'s `bias` rebase
+/// holds the reference's current frame) but not the same object: a pause is a bit the clock reads,
+/// where a speed is a *value* other code copies. `transplant_up` copies a clip's speed when it moves a
+/// one-shot to the torso overlay, and that is how the first version welded a cast pose to the upper
+/// body permanently — see [`AnimRateFreeze`] and decision 0891. Nothing here reads or writes a speed, so nothing can
+/// inherit the freeze and outlive it.
 ///
 /// **Reasserted every frame, and that is the faithful shape, not a belt-and-braces clamp.** The
 /// reference's rate lives on the bone (`+0xb0`, the factor `timebase.md` §2's clock multiplies its
 /// window by) and **arming an animation does not touch it**: op4 `0x7121a0`'s success leg writes the
 /// track index and the cursors and leaves `+0xb0` alone (only its disarm leg and `SetBoneAnimSpeed
 /// 0x712910` write it). So "rate 0 until the aura is reaped" is literally the reference's state
-/// across every re-arm in between, and re-writing it each frame is how a per-clip speed says the
-/// same thing.
+/// across every re-arm in between, and re-pausing each frame is how a per-clip flag says the same
+/// thing — including for clips armed *under* the freeze, which the reference freezes by construction.
 ///
 /// It runs **after the driver**, so this frame's arms are already in — which is the whole Ice Block
 /// observable: the cast one-shot is armed and then frozen at the frame it was on, so the caster never
-/// gets to raise their hand (`0x712910` re-bases `bias` to preserve the current frame, so rate 0
-/// holds the pose exactly where it stood).
+/// gets to raise their hand.
+///
+/// **A non-zero rate does nothing** (and only kit 1744's `8947848.0` is one — a packed grey that
+/// landed in the rate column, on Stoned / Petrification / Thadius Spawn). Our clock is f32 seconds
+/// where the reference's is integer milliseconds with a modulo, so a rate that large is noise there
+/// and different noise here; there is no honest transliteration, and inventing one would be worse than
+/// saying so. Recorded, not special-cased.
 ///
 /// What it deliberately does **not** touch: attached effect models (the ice block's own
 /// `icebarrier_state.mdx` keeps shimmering — the reference writes the unit's two models only) and
@@ -356,65 +374,59 @@ pub(crate) fn apply_aura_anim_rate(
        &AuraNodes,
        Option<&crate::entities::mount::MountChild>,
    )>,
-    mut rigs: Query<(&mut AnimationPlayer, Option<&mut AnimRateFreeze>)>,
+    mut rigs: Query<(&mut AnimationPlayer, Has<AnimRateFreeze>)>,
    mut commands: Commands,
 ) {
    for (entity, nodes, mount) in &units {
-        let rate = nodes.head_anim_rate();
+        let frozen = nodes.head_anim_rate() == Some(0.0);
        for rig in [Some(entity), mount.map(|m| m.0)].into_iter().flatten() {
-            let Ok((mut player, freeze)) = rigs.get_mut(rig) else {
+            let Ok((mut player, held)) = rigs.get_mut(rig) else {
                continue;
            };
-            match (rate, freeze) {
-                (Some(rate), Some(mut freeze)) => hold(&mut player, &mut freeze, rate),
-                (Some(rate), None) => {
-                    let mut fresh = AnimRateFreeze::default();
-                    hold(&mut player, &mut fresh, rate);
-                    trace_clock_edge("freeze", rig, rate, &fresh);
-                    commands.entity(rig).try_insert(fresh);
-                }
-                (None, Some(freeze)) => {
-                    for (node, speed) in &freeze.saved {
-                        if let Some(anim) = player.animation_mut(*node) {
-                            anim.set_speed(*speed);
-                        }
+            match (frozen, held) {
+                (true, _) => {
+                    for (_, anim) in player.playing_animations_mut() {
+                        anim.pause();
                    }
-                    trace_clock_edge("thaw", rig, 1.0, &freeze);
+                    if !held {
+                        trace_clock_edge("freeze", rig, &player);
+                        commands.entity(rig).try_insert(AnimRateFreeze);
+                    }
+                }
+                (false, true) => {
+                    // Resume everything: the set we paused is "whatever was playing", and it grew
+                    // every frame the freeze held. Nothing else on a unit rig pauses an animation
+                    // (the portrait booth does, on its own rigs), so there is no third party's pause
+                    // to trample.
+                    for (_, anim) in player.playing_animations_mut() {
+                        anim.resume();
+                    }
+                    trace_clock_edge("thaw", rig, &player);
                    commands.entity(rig).try_remove::<AnimRateFreeze>();
                }
-                (None, None) => {}
+                (false, false) => {}
            }
        }
    }
 }

-/// Take every playing animation's clock to `rate`, remembering the speed each one was running at the
-/// first time we touch it — a clip armed *during* the freeze is saved on its own first frame, so the
-/// release hands back what the driver meant rather than a blanket 1.0.
-fn hold(player: &mut AnimationPlayer, freeze: &mut AnimRateFreeze, rate: f32) {
-    for (node, anim) in player.playing_animations_mut() {
-        if !freeze.saved.iter().any(|(n, _)| n == node) {
-            freeze.saved.push((*node, anim.speed()));
-        }
-        anim.set_speed(rate);
-    }
-}
-
 /// The freeze's own instrument (`WOW_MOVE_TRACE`, tag `aur`): one line per rig at each edge, naming
-/// the clips it took over and the speeds it is holding — the difference between "the clocks stopped"
-/// and "nothing was playing anyway", which no count of *absent* anim events can tell apart.
-fn trace_clock_edge(what: &str, rig: Entity, rate: f32, freeze: &AnimRateFreeze) {
+/// the clips it took over and the speed each is *left* running at — the difference between "the clocks
+/// stopped" and "nothing was playing anyway", which no count of *absent* anim events can tell apart.
+///
+/// It prints the speeds precisely because the first version's bug was invisible without them: a `@0`
+/// on the thaw line is a clip that will never advance again.
+fn trace_clock_edge(what: &str, rig: Entity, player: &AnimationPlayer) {
    if !crate::dbg_trace::enabled_for("aur") {
        return;
    }
-    let clips: Vec<String> = freeze
-        .saved
-        .iter()
-        .map(|(n, s)| format!("{}@{s}", n.index()))
+    let clips: Vec<String> = player
+        .playing_animations()
+        .map(|(n, a)| format!("{}@{}", n.index(), a.speed()))
        .collect();
    crate::dbg_trace::line(
        "aur",
-        &format!("{what} rig={rig} rate={rate} clips=[{}]", clips.join(" ")),
+        &format!("{what} rig={rig} clips=[{}]", clips.join(" ")),
    );
 }

@@ -773,39 +773,41 @@ fn app_with_freeze() -> App {

 /// A rig mid-animation: two clips playing at the speeds the driver picked for them (a gait scaled to
 /// the mover's speed, and a cast one-shot at 1.0 — the pair Ice Block catches).
-fn rig_playing(speeds: [(u32, f32); 2]) -> AnimationPlayer {
+fn rig_playing(speeds: &[(u32, f32)]) -> AnimationPlayer {
    let mut player = AnimationPlayer::default();
    for (node, speed) in speeds {
        player
-            .play(AnimationNodeIndex::new(node as usize))
-            .set_speed(speed);
+            .play(AnimationNodeIndex::new(*node as usize))
+            .set_speed(*speed);
    }
    player
 }

-fn speed_of(app: &App, rig: Entity, node: u32) -> f32 {
-    app.world()
+/// `(paused, speed)` of one armed clip — the two things the freeze is judged on: the clock is held,
+/// and the *value* other code copies is untouched.
+fn clip(app: &App, rig: Entity, node: u32) -> (bool, f32) {
+    let a = app
+        .world()
        .entity(rig)
        .get::<AnimationPlayer>()
        .expect("the rig kept its player")
        .animation(AnimationNodeIndex::new(node as usize))
-        .expect("the clip is still armed")
-        .speed()
+        .expect("the clip is still armed");
+    (a.is_paused(), a.speed())
 }

 /// **The director's retest, at the mechanism.** Ice Block's state kit carries proc 11 at rate 0, so
 /// the unit's clocks stop dead: the run cycle it was mid-stride in and the cast one-shot that had
 /// just been armed both hold the frame they were on — which is why the caster never gets to raise
-/// their hand. Dropping the aura hands each clip back the speed the driver had given it, not a
-/// blanket 1.0.
+/// their hand. Dropping the aura lets both go again, at the speeds the driver gave them.
 #[test]
-fn ice_block_holds_the_clocks_and_the_drop_hands_them_back() {
+fn ice_block_holds_the_clocks_and_the_drop_lets_them_go() {
    let mut app = app_with_freeze();
    let unit = app
        .world_mut()
        .spawn((
            crate::net::ObjectStore(ObjectFields::from_pairs(&[(47, ICE_BLOCK), (95, 0x0E)])),
-            rig_playing([(4, 1.35), (54, 1.0)]),
+            rig_playing(&[(4, 1.35), (54, 1.0)]),
        ))
        .id();
    app.update();
@@ -814,15 +816,16 @@ fn ice_block_holds_the_clocks_and_the_drop_hands_them_back() {
    assert_eq!(n.head_anim_rate(), Some(0.0), "kit 3709's proc 11 is 0");
    // 9074175 = 0x8A75FF — the ice-blue the same kit tints the body with, beside the freeze.
    assert_eq!(n.head_tint(), Some([0x8A, 0x75, 0xFF]), "kit 3709's proc 1");
-    assert_eq!(speed_of(&app, unit, 4), 0.0, "the gait stops mid-stride");
-    assert_eq!(speed_of(&app, unit, 54), 0.0, "so does the cast one-shot");
+    assert!(clip(&app, unit, 4).0, "the gait stops mid-stride");
+    assert!(clip(&app, unit, 54).0, "so does the cast one-shot");
+    // The SPEEDS are untouched, which is the whole point: nothing that copies a clip's rate
+    // (`transplant_up` does, to move a one-shot onto the torso overlay) can inherit the freeze.
+    assert_eq!(clip(&app, unit, 4).1, 1.35);
+    assert_eq!(clip(&app, unit, 54).1, 1.0);

-    // A second frame changes nothing — the hold is idempotent, and it must not re-save the zero it
-    // wrote itself (that is what would make the release hand back 0.0 forever).
-    app.update();
-    assert_eq!(speed_of(&app, unit, 4), 0.0);
+    app.update(); // idempotent

-    // The aura leaves the slots: the reference writes the saved rates back (`0x6203e0`).
+    // The aura leaves the slots.
    app.world_mut()
        .entity_mut(unit)
        .insert(crate::net::ObjectStore(ObjectFields::from_pairs(&[(
@@ -836,41 +839,53 @@ fn ice_block_holds_the_clocks_and_the_drop_hands_them_back() {
        .unwrap()
        .rate
        .is_empty());
-    assert_eq!(
-        speed_of(&app, unit, 4),
-        1.35,
-        "the gait's own rate, restored"
-    );
-    assert_eq!(speed_of(&app, unit, 54), 1.0);
+    assert!(!clip(&app, unit, 4).0, "running again");
+    assert!(!clip(&app, unit, 54).0);
+    assert_eq!(clip(&app, unit, 4).1, 1.35, "at its own rate");
    assert!(
        app.world().entity(unit).get::<AnimRateFreeze>().is_none(),
-        "the release drops the save with it"
+        "the release drops the marker with it"
    );
 }

-/// A clip armed *during* the freeze is saved on the frame we first touch it, so the release hands
-/// back what the driver meant for it rather than the frozen 0.
+/// **The bug the first shipped version had, pinned so it cannot come back.** A one-shot moved to the
+/// torso overlay under the freeze (`transplant_up` copies the source clip's speed and seek) used to
+/// be saved at the frozen `0`, restored to `0`, and left welded to the upper body for the rest of the
+/// session — surviving every later cast, because a clip that never advances never *finishes* and so
+/// never releases the overlay. Pausing carries no value to copy: the transplant inherits the driver's
+/// real speed, and the thaw sets it running.
 #[test]
-fn a_clip_armed_under_the_freeze_is_saved_before_it_is_held() {
+fn a_clip_armed_under_the_freeze_comes_back_at_its_own_speed() {
    let mut app = app_with_freeze();
    let unit = app
        .world_mut()
        .spawn((
            crate::net::ObjectStore(ObjectFields::from_pairs(&[(47, ICE_BLOCK), (95, 0x0E)])),
-            rig_playing([(4, 1.35), (54, 1.0)]),
+            rig_playing(&[(4, 1.35), (54, 1.0)]),
        ))
        .id();
    app.update();

-    // The driver arms Stand at 1.0 a frame later (the root wiped the direction bits).
-    app.world_mut()
-        .entity_mut(unit)
-        .get_mut::<AnimationPlayer>()
+    // `transplant_up`'s own shape: read the frozen source clip, arm the torso twin from it.
+    let speed = clip(&app, unit, 54).1;
+    assert_eq!(speed, 1.0, "the source's speed is not the freeze");
+    let seek = app
+        .world()
+        .entity(unit)
+        .get::<AnimationPlayer>()
        .unwrap()
-        .play(AnimationNodeIndex::new(0))
-        .set_speed(1.0);
+        .animation(AnimationNodeIndex::new(54))
+        .unwrap()
+        .seek_time();
+    {
+        let mut ent = app.world_mut().entity_mut(unit);
+        let mut player = ent.get_mut::<AnimationPlayer>().unwrap();
+        let upper = player.play(AnimationNodeIndex::new(156));
+        upper.seek_to(seek);
+        upper.set_speed(speed);
+    }
    app.update();
-    assert_eq!(speed_of(&app, unit, 0), 0.0, "held the frame it arrived on");
+    assert!(clip(&app, unit, 156).0, "held on arrival");

    app.world_mut()
        .entity_mut(unit)
@@ -878,7 +893,12 @@ fn a_clip_armed_under_the_freeze_is_saved_before_it_is_held() {
            95, 0,
        )])));
    app.update();
-    assert_eq!(speed_of(&app, unit, 0), 1.0);
+    assert!(!clip(&app, unit, 156).0);
+    assert_eq!(
+        clip(&app, unit, 156).1,
+        1.0,
+        "the overlay runs out and releases instead of welding a cast pose on"
+    );
 }

 /// The mount comes with the rider: `0x6201d0` writes `[unit+0xdc]` before it writes `[unit+0xd8]`,
@@ -888,20 +908,20 @@ fn the_freeze_reaches_the_mount_body() {
    let mut app = app_with_freeze();
    let mount = app
        .world_mut()
-        .spawn(rig_playing([(5, 2.0), (91, 1.0)]))
+        .spawn(rig_playing(&[(5, 2.0), (91, 1.0)]))
        .id();
    let unit = app
        .world_mut()
        .spawn((
            crate::net::ObjectStore(ObjectFields::from_pairs(&[(47, ICE_BLOCK), (95, 0x0E)])),
-            rig_playing([(91, 1.0), (54, 1.0)]),
+            rig_playing(&[(91, 1.0), (54, 1.0)]),
            crate::entities::mount::MountChild(mount),
        ))
        .id();
    app.update();

-    assert_eq!(speed_of(&app, unit, 91), 0.0, "the rider");
-    assert_eq!(speed_of(&app, mount, 5), 0.0, "and the mount under them");
+    assert!(clip(&app, unit, 91).0, "the rider");
+    assert!(clip(&app, mount, 5).0, "and the mount under them");

    app.world_mut()
        .entity_mut(unit)
@@ -909,7 +929,8 @@ fn the_freeze_reaches_the_mount_body() {
            95, 0,
        )])));
    app.update();
-    assert_eq!(speed_of(&app, mount, 5), 2.0);
+    assert!(!clip(&app, mount, 5).0);
+    assert_eq!(clip(&app, mount, 5).1, 2.0);
 }

 /// An aura with no proc 11 — Stealth — leaves the clocks entirely alone. The freeze is the kit's,
@@ -921,10 +942,27 @@ fn an_aura_without_the_proc_never_touches_a_clock() {
        .world_mut()
        .spawn((
            crate::net::ObjectStore(ObjectFields::from_pairs(&[(47, STEALTH), (95, 0x0E)])),
-            rig_playing([(4, 1.35), (54, 1.0)]),
+            rig_playing(&[(4, 1.35), (54, 1.0)]),
        ))
        .id();
    app.update();
-    assert_eq!(speed_of(&app, unit, 4), 1.35);
+    assert!(!clip(&app, unit, 4).0);
    assert!(app.world().entity(unit).get::<AnimRateFreeze>().is_none());
 }
+
+/// Kit 1744's `8947848.0` is a rate our f32-seconds clock cannot express (the reference's is integer
+/// ms with a modulo, where it is noise rather than a freeze). It installs a node — the data is read,
+/// not dropped — and the apply leaves the clocks alone rather than inventing a look for it.
+#[test]
+fn a_non_zero_rate_installs_a_node_and_holds_nothing() {
+    let mut n = AuraNodes::new(1.0);
+    install(&mut n, 17624, &[AuraNode::AnimRate(8_947_848.0)], 0.0);
+    assert_eq!(n.head_anim_rate(), Some(8_947_848.0));
+
+    let mut app = app_with_freeze();
+    let unit = app.world_mut().spawn(rig_playing(&[(4, 1.35)])).id();
+    app.world_mut().entity_mut(unit).insert(n);
+    app.update();
+    assert!(!clip(&app, unit, 4).0);
+    assert_eq!(clip(&app, unit, 4).1, 1.35);
+}
@@ -521,6 +521,14 @@ pub(crate) struct AnimDriver {
    /// The gait id currently targeted while in [`Mode::Gait`] (so we cross-fade only on a change). `None`
    /// forces a fresh selection (first eval, or after a Special state exits).
    gait: Option<u16>,
+    /// The movement flags the **base slot** was last armed for — the reference has no per-one-shot
+    /// latch, so "the movement state changed" means *since the base last took a request*, never
+    /// "since this one-shot started" (decision 0894). The distinction is invisible until a one-shot
+    /// and a movement change land on the SAME frame: Ice Block's root wipes the direction bits in
+    /// the very frame the cast one-shot arrives, so an arm-time comparison sees no edge at all and
+    /// the cast keeps bone 0 — where the reference's next base request overwrites it and leaves the
+    /// character neutral.
+    gait_flags: u32,
    /// The unit's **client-side sheath state** — the mirror of the client's committed CUR cache
    /// (`[unit+0xd40]`, decision 0080): what the weapon placement renders (absent a
    /// [`VisualSheath`] ceremony pin) and what the setter/reconcile test against. Seeded from
@@ -661,6 +669,7 @@ impl Default for AnimDriver {
        Self {
            mode: Mode::Gait,
            gait: None,
+            gait_flags: 0,
            sheath_cur: None,
            sheath_byte: None,
            sheath_swap: None,
@@ -845,11 +845,7 @@ pub(super) fn drive_animations(
                if let Some((c, repeat)) = picked {
                    play_clip(&mut tr, &mut player, c, repeat, 1.0);
                }
-                drv.mode = Mode::Swing {
-                    id,
-                    flags: mv.flags,
-                    under: special,
-                };
+                drv.mode = Mode::Swing { id, under: special };
                drv.gait = None;
                base_played = true;
                played_oneshot = Some(id);
@@ -1044,7 +1040,7 @@ pub(super) fn drive_animations(
                    drv.gait = None; // recompute a fresh gait next frame
                }
            }
-            Mode::Swing { id, flags, under } => {
+            Mode::Swing { id, under } => {
                if special != under {
                    // The state this one-shot replaced CHANGED — the client's next event play
                    // supersedes it: a fresh jump/pose entry (`None → Some`), the FALLINGFAR
@@ -1125,25 +1121,47 @@ pub(super) fn drive_animations(
                        );
                        drv.mode = Mode::Looping(sp);
                    }
-                } else if oneshot_finished(&player, anims, id, catalog) || mv.flags != flags {
-                    // Finished — or a movement-flag change: the client's locomotion re-arm lands
-                    // on the change and blindly overwrites bone 0, one-shot or not (the same
-                    // re-arm decision 0280 named for the un-finishable looping kit clip, and
+                } else if oneshot_finished(&player, anims, id, catalog)
+                    || mv.flags != drv.gait_flags
+                {
+                    // Finished — or a movement-flag change: the client's base re-arm lands on the
+                    // change and blindly overwrites bone 0, one-shot or not (the same re-arm
+                    // decision 0280 named for the un-finishable looping kit clip, and
                    // Mode::Land's flag-change re-pick). Holding the clip out instead slides the
                    // post-shot runner over the ground on straight legs (director-observed vs
                    // ref). An EDGE, not a level — the split-boneless masked fallback enters
                    // here already moving, and steady flags must let that clip play out.
-                    if mv.flags != flags {
+                    //
+                    // Against `drv.gait_flags` — what the **base** was last armed for — not this
+                    // one-shot's own arm-time `flags` (decision 0894). The reference keeps no
+                    // per-one-shot latch: a movement-state change requests the base and bone 0
+                    // takes it, whenever the one-shot happened to start. Ice Block is the case that
+                    // separates them — its root wipes the direction bits in the SAME frame the cast
+                    // one-shot arrives, so the arm-time compare sees no edge ever and the cast held
+                    // bone 0 for the whole block; against the base's flags the edge is still there,
+                    // Stand overwrites the cast, and the character is neutral when the freeze lands.
+                    if mv.flags != drv.gait_flags {
                        // The locomotion re-arm is a normal PlayAnimation — the deferred-combat
                        // cache clears with it (decision 0406; a finished clip instead had its
                        // cache consumed by the injection above, before this machine ran).
                        drv.deferred = None;
-                        // …and it is a LOCOMOTION request, so a still-playing CAST/COMBAT clip
-                        // transplants up to the key-bone instead of being overwritten (decision
-                        // 0878). A *finished* clip has nothing to move: the client's descriptor
-                        // probe reports a completed slot as id −1 (`0x5fe1f0` reads the completion
-                        // latch, not the armed record), so the transplant predicates never see it.
-                        transplant_up(&mut drv, &mut player, &tr, anims, entity, id);
+                        // …and **if** the re-arm resolves to a LOCOMOTION id, a still-playing
+                        // CAST/COMBAT clip transplants up to the key-bone instead of being
+                        // overwritten (decision 0878; the client's gate is `0x5fee80` on the
+                        // *requested* id, `0x5fe912`). A *finished* clip has nothing to move: the
+                        // client's descriptor probe reports a completed slot as id −1 (`0x5fe1f0`
+                        // reads the completion latch, not the armed record), so the transplant
+                        // predicates never see it.
+                        //
+                        // The gate was missing here, and Ice Block is what it costs (decision
+                        // 0894): a stun's root wipes the direction bits, so the flag change re-arms
+                        // to **Stand(0)** — not locomotion — and the reference *overwrites* the
+                        // cast on bone 0, leaving the character fully neutral for the freeze to
+                        // catch. Transplanting unconditionally moved it to the torso instead and
+                        // froze an arm out.
+                        if select::gait_is_locomotion(&mv, walk) {
+                            transplant_up(&mut drv, &mut player, &tr, anims, entity, id);
+                        }
                    }
                    drv.mode = Mode::Gait;
                    drv.gait = None; // recompute a fresh gait next frame
@@ -1377,6 +1395,11 @@ pub(super) fn drive_animations(
                    } else {
                        drv.gait = Some(target); // no clip (bind pose) — record the target so we don't churn
                    }
+                    // The base is now arm-consistent with this movement state — every branch above
+                    // leaves `drv.gait == Some(target)`. A one-shot that displaces it later reads
+                    // this, not its own arm-time flags, to know whether the movement state has
+                    // moved on since (decision 0894).
+                    drv.gait_flags = mv.flags;
                    // Leaving the ranged idle (moving, sheath change, cancel) drops the hold —
                    // the next shooting session opens with a fresh nock.
                    if ranged_load.is_none() {
@@ -649,7 +649,6 @@ fn same_frame_collision_fast_paths_the_second_combat_clip() {
        drv(&app, spin_last).mode,
        super::super::select::Mode::Swing {
            id: 16,
-            flags: 0,
            under: None,
        },
        "the first arrival holds the body"
@@ -662,7 +661,6 @@ fn same_frame_collision_fast_paths_the_second_combat_clip() {
        drv(&app, swing_last).mode,
        super::super::select::Mode::Swing {
            id: 57,
-            flags: 0,
            under: None,
        },
        "the spin holds the body through the later swing"
@@ -712,7 +710,6 @@ fn deferred_combat_clip_plays_once_the_body_frees() {
        drv.mode,
        super::super::select::Mode::Swing {
            id: 16,
-            flags: 0,
            under: None,
        },
        "the parked swing armed"
@@ -779,7 +776,6 @@ fn a_movement_flag_change_cuts_a_full_body_oneshot_immediately() {
        mode(&app),
        super::super::select::Mode::Swing {
            id: 16,
-            flags: 0,
            under: None,
        },
        "standing swing holds the base track"
@@ -1331,7 +1327,6 @@ fn a_jump_in_place_cast_replaces_the_hang_and_survives_the_arc() {
        drv(&app, unit).mode,
        super::super::select::Mode::Swing {
            id: 54,
-            flags: move_flags::FALLING,
            under: Some(super::super::select::Special::Jump),
        },
        "the cast replaces the hang full-body — never dropped, never masked"
@@ -2114,3 +2109,100 @@ fn a_melee_to_ranged_toggle_stows_both_hands_before_it_reaches_for_the_bow() {
        "the ceremony must end with the pin dropped, leaving the committed state (saw {seen:?})"
    );
 }
+
+/// **Ice Block, at the mechanism (decision 0894).** The stun's root wipes the direction bits in the
+/// SAME frame the cast one-shot arrives, so the one-shot's own arm-time flags already read ROOT and
+/// an arm-time comparison sees no movement edge — ever. The base's flags still hold the run it was
+/// armed for, so the edge is there, the re-arm resolves to **Stand(0)** (not locomotion → no
+/// transplant), and Stand overwrites the cast on bone 0: the character is fully neutral, with
+/// nothing on the torso, by the time the freeze catches it. Before this the cast held bone 0 for the
+/// whole block, or rode up to the torso and froze an arm out.
+#[test]
+fn a_root_landing_with_the_cast_returns_the_body_to_neutral() {
+    let mut app = app();
+    let unit = jumper(&mut app);
+    app.world_mut().entity_mut(unit).insert(MovementState {
+        flags: move_flags::FORWARD,
+        speed: 7.0,
+        ..Default::default()
+    });
+    app.update();
+    let gait = |app: &App| app.world().entity(unit).get::<AnimDriver>().unwrap().gait;
+    let mode = |app: &App| app.world().entity(unit).get::<AnimDriver>().unwrap().mode;
+    let overlay = |app: &App| {
+        app.world()
+            .entity(unit)
+            .get::<AnimDriver>()
+            .unwrap()
+            .overlay
+            .map(|o| o.id)
+    };
+    assert_eq!(gait(&app), Some(5), "running");
+
+    // One frame: the root lands (direction bits wiped) AND the cast one-shot arrives.
+    app.world_mut().entity_mut(unit).insert(MovementState {
+        flags: move_flags::ROOT,
+        ..Default::default()
+    });
+    app.world_mut().write_message(EmoteAnim {
+        entity: unit,
+        anim_id: 54,
+        seq: 1,
+    });
+    app.update();
+    // The cast takes bone 0 and the base re-arm displaces it inside the same frame — the reference's
+    // one-slot last-writer-wins, which is exactly what the base's flags (still FORWARD) unlock.
+    assert_eq!(
+        mode(&app),
+        super::super::select::Mode::Gait,
+        "the base request wins bone 0 back"
+    );
+    assert!(
+        overlay(&app).is_none(),
+        "no torso transplant: Stand(0) is not a locomotion request"
+    );
+    app.update();
+    assert_eq!(gait(&app), Some(0), "fully neutral standing");
+    assert!(overlay(&app).is_none(), "and nothing left on the torso");
+}
+
+/// The control that keeps 0878 honest: the transplant still fires when the re-arm really *is*
+/// locomotion. A cast fired standing, then the player runs — the legs take Run and the cast keeps
+/// going on the torso, which is the whole of R-B.
+#[test]
+fn a_run_starting_under_a_cast_still_transplants_it_up() {
+    let mut app = app();
+    let unit = jumper(&mut app);
+    app.update(); // settle: Stand
+    app.world_mut().write_message(EmoteAnim {
+        entity: unit,
+        anim_id: 54,
+        seq: 1,
+    });
+    app.update();
+    let mode = |app: &App| app.world().entity(unit).get::<AnimDriver>().unwrap().mode;
+    let overlay = |app: &App| {
+        app.world()
+            .entity(unit)
+            .get::<AnimDriver>()
+            .unwrap()
+            .overlay
+            .map(|o| o.id)
+    };
+    assert!(matches!(
+        mode(&app),
+        super::super::select::Mode::Swing { id: 54, .. }
+    ));
+
+    app.world_mut().entity_mut(unit).insert(MovementState {
+        flags: move_flags::FORWARD,
+        speed: 7.0,
+        ..Default::default()
+    });
+    app.update();
+    assert_eq!(
+        overlay(&app),
+        Some(54),
+        "a locomotion re-arm moves the cast to the torso instead of cutting it"
+    );
+}
@@ -405,18 +405,16 @@ pub(super) enum Mode {
    /// *edge* exits — the FALLINGFAR latch's Fall (played ONCE, at the latch edge
    /// `0x61a820` — 0864's per-tick re-assert was §5-refuted, decision 0868; a clip armed
    /// after the latch holds like any other) and the `0x602c60` land pick at touchdown. Grounded
-    /// (`under == None`) it returns to Gait when it finishes; a new one restarts it — and, like
-    /// [`Mode::Land`], any movement-flag change from `flags` (captured at play time) re-picks
-    /// immediately: the client's locomotion re-arm overwrites bone 0 blindly on the change (the
-    /// same re-arm decision 0280 named), so a shot fired standing yields to the run the instant
-    /// the player moves instead of sliding the clip out. The *masked* route never enters here —
-    /// it plays as an overlay beside `Mode`, leaving the base machine (this enum) running
-    /// underneath.
-    Swing {
-        id: u16,
-        flags: u32,
-        under: Option<Special>,
-    },
+    /// (`under == None`) it returns to Gait when it finishes; a new one restarts it — and any
+    /// movement-flag change since the **base** was last armed
+    /// ([`AnimDriver::gait_flags`](crate::creature_anim::AnimDriver)) re-picks immediately: the
+    /// client's base re-arm overwrites bone 0 blindly on the change (the same re-arm decision 0280
+    /// named), so a shot fired standing yields to the run the instant the player moves instead of
+    /// sliding the clip out. It carried its own arm-time flags until 0894, which made the edge
+    /// invisible whenever the one-shot and the movement change landed on the same frame. The
+    /// *masked* route never enters here — it plays as an overlay beside `Mode`, leaving the base
+    /// machine (this enum) running underneath.
+    Swing { id: u16, under: Option<Special> },
 }

 /// The gait an idle/moving unit should play in [`Mode::Gait`], most-specific id first with fallbacks so a
@@ -975,6 +973,25 @@ pub(super) fn is_locomotion(id: u16) -> bool {
    RATE_SCALED.contains(&id)
 }

+/// Whether the gait this movement state resolves to is a **locomotion** id — i.e. whether a
+/// movement-flag re-arm is the transplant's `0x5fee80` request or an ordinary bone-0 overwrite.
+///
+/// The re-arm at a flag change has no id in hand yet (the gait is recomputed the following frame),
+/// so this asks [`gait_candidates`] what it *will* pick. `ready`/`ranged_load` are passed `None`
+/// deliberately and it changes no answer: both are **standing** idles that locomotion outranks in
+/// the cascade, so a moving unit picks its locomotion id whether or not they are set, and a still
+/// one picks something that is not locomotion either way.
+///
+/// This is the gate Ice Block turns on. A stun's root wipes the direction bits, the flag change
+/// re-arms, and the pick is **Stand(0)** — *not* locomotion — so the cast one-shot on bone 0 is
+/// **overwritten** rather than transplanted onto the torso: the character goes fully neutral before
+/// the freeze catches it (decision 0894, the director's reading of the reference client).
+pub(super) fn gait_is_locomotion(state: &MovementState, walk_speed: f32) -> bool {
+    gait_candidates(state, walk_speed, None, None)
+        .first()
+        .is_some_and(|id| is_locomotion(*id))
+}
+
 /// The playback rate for a clip given the unit's live speed (wow-5875-re `0x5fe2f0`): `speed / moveSpeed`
 /// for a rate-scaled locomotion clip, else 1×. (The client also divides by the model's scale magnitude,
 /// 1.0 for an unscaled character.) Guarded on `moveSpeed > 0`.
@@ -249,6 +249,9 @@ mod tests {
    use super::*;

    use super::test_template as info;
+    use benilla_protocol::messages::{
+        ITEM_DYNFLAG_UNLOCKED, ITEM_DYNFLAG_WRAPPED, ITEM_FLAG_LOOTABLE, ITEM_FLAG_WRAPPER,
+    };
    use crossbeam_channel::TryRecvError;

    fn commands() -> (NetCommands, crossbeam_channel::Receiver<ClientCommand>) {
@@ -291,6 +294,35 @@ mod tests {
        assert!(items.take_fresh().is_empty(), "a drain empties the queue");
    }

+    /// The one predicate both halves of the right-click-to-open affordance ask
+    /// (`ItemInfo::openable`): the tooltip's green line and the `CMSG_OPEN_ITEM` fork. Its whole
+    /// subtlety is the two sub-gates — a `LockID` template stays shut until the INSTANCE carries
+    /// UNLOCKED, and a wrapper template needs the instance's WRAPPED bit — so a template-only
+    /// view (instance flags 0) can never be openable.
+    #[test]
+    fn openable_is_the_lootable_bit_behind_its_lock_and_gift_sub_gates() {
+        let plain = |flags: u32, lock_id: u32| {
+            let mut t = info("Small Barnacled Clam");
+            t.flags = flags;
+            t.lock_id = lock_id;
+            t
+        };
+
+        // The clam: LOOTABLE, no lock — openable outright, instance flags irrelevant.
+        assert!(plain(ITEM_FLAG_LOOTABLE, 0).openable(0));
+        // An ordinary item is never openable, however its instance is flagged.
+        assert!(!plain(0, 0).openable(ITEM_DYNFLAG_UNLOCKED | ITEM_DYNFLAG_WRAPPED));
+        // A junkbox: LOOTABLE but locked — shut until the instance says UNLOCKED.
+        assert!(!plain(ITEM_FLAG_LOOTABLE, 7).openable(0));
+        assert!(plain(ITEM_FLAG_LOOTABLE, 7).openable(ITEM_DYNFLAG_UNLOCKED));
+        // Gift wrap: the WRAPPER template opens only while the instance is still WRAPPED.
+        assert!(!plain(ITEM_FLAG_WRAPPER, 0).openable(0));
+        assert!(plain(ITEM_FLAG_WRAPPER, 0).openable(ITEM_DYNFLAG_WRAPPED));
+        // The wrapped arm ignores the lock gate entirely — it is the sibling `||`, not a nested
+        // case: a wrapped gift of a locked box opens as the gift.
+        assert!(plain(ITEM_FLAG_WRAPPER | ITEM_FLAG_LOOTABLE, 7).openable(ITEM_DYNFLAG_WRAPPED));
+    }
+
    #[test]
    fn negative_template_answer_is_cached() {
        let (cmds, rx) = commands();
@@ -600,6 +600,12 @@ pub(crate) enum ClientCommand {
        spell_index: u8,
        go_target: Option<u64>,
    },
+    /// Open a bag item (`CMSG_OPEN_ITEM`, same bag addressing) — the drain's fork for an
+    /// *openable* click (`ItemInfo::openable`): a clam, an unlocked lockbox, a wrapped gift. The
+    /// server answers `SMSG_LOOT_RESPONSE` on the **item's own guid**, so the ordinary loot feed
+    /// opens a window over a thing in the bag; a wrapped gift instead swaps entry in place and
+    /// sends no window. Refusals arrive as `InventoryFailure` (still locked, dead, flying).
+    OpenItem { bag_index: u8, slot: u8 },
    /// Equip a bag item (`CMSG_AUTOEQUIP_ITEM`, same bag addressing) — the drain's fork for an
    /// *equippable* click, mirroring the real client's equip-vs-use decision. Refusals come back
    /// as `InventoryFailure` events onto the UI error line.
@@ -663,6 +663,7 @@ fn writer_loop(
                        spell_index,
                        go_target,
                    } => w.use_item(bag_index, slot, spell_index, go_target),
+                    ClientCommand::OpenItem { bag_index, slot } => w.open_item(bag_index, slot),
                    ClientCommand::AutoEquipItem { bag_index, slot } => {
                        w.auto_equip_item(bag_index, slot)
                    }
@@ -31,6 +31,8 @@ use crate::ui_script::PointerOverUi;

 mod arc;
 mod camera;
+mod follow;
+
 mod gait;
 // The land-here affordance (free-fly's other half). `pub(crate)` for its `LandHere` message, which
 // the debug panel's button writes.
@@ -54,11 +56,14 @@ mod wire_in;
 // render-alpha field, so the unit-lane material-alpha compose (`entities::apply_unit_mat_alpha`)
 // orders itself before it and lets that documented override stand.
 pub(crate) use camera::apply_self_model_fade;
+// `/follow` (decision 0890): chat asks with the message, `crate::target` resolves the subject into
+// the state, and this module owns the motion.
 use camera::{
    apply_zoom_scroll, run_look_session, seat_camera, CameraProbe, FlyCam, LookButton,
    CAM_COLLISION_RADIUS, CAM_DIST_DEFAULT, CAM_PIVOT_FALLBACK,
 };
 pub(crate) use camera::{head_height, CameraControl, CameraPivot, WorldCamera, CAM_NEAR};
+pub(crate) use follow::{FollowRequest, FollowState};
 // The shared avatar state + movement constants live in [`state`]; the private re-imports below are
 // what lets this module and the concern modules beside it keep naming them `super::X` unchanged.
 use state::{
@@ -125,6 +130,7 @@ impl Plugin for PlayerPlugin {
        if let Some(cam) = probe_cam::from_env() {
            app.insert_resource(cam);
        }
+        follow::plugin(app);
        app.add_systems(Startup, setup::setup_player.after(AssetSet::Open))
            // The world camera renders only when the world can be seen (decision 0540): in world,
            // or under the opaque loading screen (whose covered render is what compiles the
@@ -204,6 +210,17 @@ impl Plugin for PlayerPlugin {
                    .in_set(WorldStage::Input)
                    .before(control),
            )
+            // `/follow` (decision 0890): steer the facing and decide this tick's synthesized forward
+            // input immediately BEFORE the controller, which folds the flag into its forward axis.
+            // The player's own turn input therefore runs after us and wins, which is exactly what
+            // makes the turn-away cancel reachable.
+            .add_systems(
+                Update,
+                follow::steer_follow
+                    .in_set(WorldStage::Input)
+                    .before(control)
+                    .run_if(in_state(crate::char_select::ClientState::InWorld)),
+            )
            // The self-avatar zoom-in fade rides the same `MeshTag`/material channel as the interior
            // classifier + the appear/despawn fades, so it must run *after* both to win the frame while
            // fading (and yield to them otherwise). It also writes `Visibility` (the first-person
@@ -625,8 +642,13 @@ fn control(
        // and the bit survives — and the client emits MSG_MOVE_STOP with S still held. The other order
        // (autorun, then S) destroys the bit at key-down and walks you backward. Same two keys, two
        // outcomes; that asymmetry is the whole shape of the feature.
+        // `/follow` enters as the FORWARD term, not a fifth source (decision 0890): the reference's
+        // follow pushes the very same move-forward bit `0x100000` the W key does, through the same
+        // setter, so it nets against a held S and diagonals with a strafe exactly like a held W. It
+        // rides the HELD state and not the key-DOWN edge, so it never trips the autorun cancel set
+        // above — which is right: synthesized input is not a keypress.
        let fwd_axis = state::forward_axis(
-            keys_pressed(KeyCode::KeyW),
+            keys_pressed(KeyCode::KeyW) || player.follow_forward,
            keys_pressed(KeyCode::KeyS),
            both_buttons,
            autorun,
@@ -0,0 +1,342 @@
+//! `/follow` — the auto-follow movement mode (decision 0890).
+//!
+//! The finding that shapes this module: **follow synthesizes keyboard input.** It owns no
+//! translation of its own. Every state change in the reference funnels through the movement
+//! singleton's setter (`0x60e790` sets the move-forward bit `0x100000`, `0x60e7f0` clears it) — the
+//! same setter the MoveForward keybinding drives, one of ~40 stubs in the keybinding command table
+//! at `0x513de0`-`0x514273`. So follow is "hold W for me, and steer", and the right implementation
+//! reuses the controller wholesale rather than growing a second mover (wow-re
+//! `object-layer/scratch/targeting-by-name.md` PART D, §5-cross-checked).
+//!
+//! It also sends **nothing on the wire** — corroborated against vmangos, which has no follow opcode
+//! in 1.12.1. The server only ever sees the ordinary movement stream our synthesized input produces.
+//!
+//! ## The law
+//!
+//! - **Facing is steered, never snapped** (`0x6103d0`/`0x6108xx`): turn toward the followee at
+//!   [`TURN_RATE`] = π rad/s (180°/s), clamped to `min(remaining, rate × elapsed)`, and stop turning
+//!   inside a [`TURN_DEADZONE`] of 0.001 rad. That rate limit is why follow *reads* like a character
+//!   running rather than a camera snapping.
+//! - **Beeline, re-aimed every tick** (`0x610e40`): the guid and both positions are re-resolved every
+//!   tick and nothing is cached. No path, no spline — `0x670630` is a *getter* over the click-to-move
+//!   order, which follow reads and never writes.
+//! - **It never reads the followee's speed** (VERIFIED negative, whole-graph scan). You move at
+//!   whatever a held forward key yields for *you*, which is why follow falls behind a faster runner.
+//! - **A hysteresis band, not a stop distance** — see [`should_move`]. Arrive is inclusive, resume is
+//!   one-directional. The gap between the two is why follow visibly starts and stops rather than
+//!   juddering on a single threshold.
+//!
+//! ## The one INFERRED rule
+//!
+//! The reference cancels follow when the player turns ~180° away (band 160°-220°), but wow-re could
+//! not trace *what that angle is measured against* — it is explicitly flagged unresolved. Measuring
+//! it against the live bearing alone cannot be right: starting a follow while facing away from the
+//! followee would then cancel on the first tick, before the steer had a chance to turn us.
+//!
+//! So this module latches [`FollowState::aligned`] the first time the facing comes inside the band
+//! and only cancels *after* that. Observable behaviour: starting a follow facing any direction works
+//! and turns you around; deliberately turning around once you are following drops it. That matches
+//! what the band is plainly *for*, and is marked INFERRED until the reference angle is pinned.
+
+use std::f32::consts::PI;
+
+use bevy::prelude::*;
+
+use crate::net::GuidIndex;
+
+use super::state::{MoveSpeed, Player};
+
+/// Follow's facing-turn rate, rad/s — `0xc4d93c`, seeded `0x40490fdb` = π at `0x6111f9`. Its own
+/// constant, distinct from the keyboard turn rate.
+const TURN_RATE: f32 = PI;
+
+/// Inside this many radians of the bearing, follow stops turning (`0x6108bb` → `0x60e920`).
+const TURN_DEADZONE: f32 = 0.001;
+
+/// The speed the distance thresholds are normalised by — `.rdata 0x80c4d0` = 7.0, which is also
+/// vanilla's base run speed, so at normal speed the band is exactly 3.0 / 4.5 yd.
+const SPEED_NORM: f32 = 7.0;
+
+/// The base stop distance — `.rdata 0x80c4c0` = 3.0, a compile-time constant (its three writers are
+/// CRT dynamic initializers behind `_initterm`), **not** a cvar.
+const STOP_DISTANCE: f32 = 3.0;
+
+/// Resume is 1.5× the stop distance, and its speed scale is floored at 1.0 — so the band never
+/// closes below 3.0 / 4.5 yd however slowly you are moving.
+const RESUME_FACTOR: f32 = 1.5;
+
+/// How far off the bearing the facing must sit for a deliberate turn to drop the follow. The
+/// reference's band is 160°-220°; wrapped into `[0, π]` that is "at least 160° off".
+const CANCEL_ANGLE: f32 = 160.0 * PI / 180.0;
+
+/// Who we are following, and the hysteresis latch — the reference's followed-guid pair `0xc4d980`
+/// (armed `0x6111c9`, cleared `0x60fc1e`) plus the state its band implies.
+#[derive(Resource, Default)]
+pub(crate) struct FollowState {
+    /// The followee's guid, or `None` when not following.
+    pub(crate) guid: Option<u64>,
+    /// Whether the synthesized forward input is currently held. The band's latch: which threshold
+    /// applies depends on which side we are already on.
+    moving: bool,
+    /// Whether the facing has come inside [`CANCEL_ANGLE`] at least once since the follow began —
+    /// the guard on the INFERRED turn-away cancel (see the module header).
+    aligned: bool,
+    /// `WOW_FOLLOW_TRACE` bookkeeping: when the last trace line went out, and where we were.
+    /// `None` until a tick has seeded it.
+    traced_at: f64,
+    traced_pos: Option<Vec3>,
+}
+
+impl FollowState {
+    /// Begin following `guid`, from a clean band.
+    pub(crate) fn start(&mut self, guid: u64) {
+        self.guid = Some(guid);
+        self.moving = false;
+        self.aligned = false;
+        self.traced_at = 0.0;
+        self.traced_pos = None;
+    }
+
+    /// Stop following. Returns whether we actually were.
+    pub(crate) fn stop(&mut self) -> bool {
+        self.moving = false;
+        self.aligned = false;
+        self.guid.take().is_some()
+    }
+}
+
+/// `/follow [name]` — start following. `None` is the bare form (`FollowUnit("target")`), whose
+/// subject is the current selection; a name resolves **players only** and additionally through the
+/// reference's filter mode 2 (`CanAssist` + alive), which is [`crate::target`]'s to apply.
+#[derive(bevy::ecs::message::Message, Clone, Debug)]
+pub(crate) struct FollowRequest {
+    pub(crate) name: Option<String>,
+}
+
+/// The distance at which follow **arrives** and lets go of the key (`0x610ad2`-`0x610b1b`):
+/// `(speed / 7.0) × 3.0`, tested inclusively (`test ah,0x41; jp` — the `<=` edge).
+fn arrive_distance(speed: f32) -> f32 {
+    (speed / SPEED_NORM) * STOP_DISTANCE
+}
+
+/// The distance at which a stopped follow **resumes** (`0x610bc4`-`0x610c2b`):
+/// `3.0 × 1.5 × max(speed / 7.0, 1.0)` = 4.5 yd at normal run speed. Only consulted while stopped.
+fn resume_distance(speed: f32) -> f32 {
+    STOP_DISTANCE * RESUME_FACTOR * (speed / SPEED_NORM).max(1.0)
+}
+
+/// Should the synthesized forward key be held this tick? The band, as a pure function of which side
+/// we are already on — the whole reason follow starts and stops instead of juddering.
+fn should_move(was_moving: bool, distance: f32, speed: f32) -> bool {
+    if was_moving {
+        // Arrive is inclusive, so we keep going only while strictly beyond it.
+        distance > arrive_distance(speed)
+    } else {
+        distance >= resume_distance(speed)
+    }
+}
+
+/// Wrap an angle into `(-π, π]`.
+fn wrap_pi(a: f32) -> f32 {
+    let t = std::f32::consts::TAU;
+    let x = (a + PI).rem_euclid(t);
+    x - PI
+}
+
+/// The `face_yaw` that points at a horizontal delta in **Bevy** space.
+///
+/// Derived from the controller's own forward vector rather than from a coordinate convention:
+/// `control` builds `move_fwd = Quat::from_rotation_y(face_yaw) * NEG_Z`, which expands to
+/// `(-sin y, 0, -cos y)`. Solving `(-sin y, -cos y) ∝ (dx, dz)` gives `y = atan2(-dx, -dz)`. Tied to
+/// the expression it must agree with, so it cannot drift out of sign with it.
+fn bearing_to(delta: Vec3) -> f32 {
+    (-delta.x).atan2(-delta.z)
+}
+
+/// Turn `face` toward `bearing` by at most this tick's budget, or leave it alone inside the
+/// deadzone. The reference's `min(remaining, rate × elapsed)` clamp.
+fn steer(face: f32, bearing: f32, dt: f32) -> f32 {
+    let remaining = wrap_pi(bearing - face);
+    if remaining.abs() <= TURN_DEADZONE {
+        return face;
+    }
+    let budget = TURN_RATE * dt;
+    face + remaining.clamp(-budget, budget)
+}
+
+/// Drive the follow: re-resolve the followee, steer the facing, and decide whether the synthesized
+/// forward input is held this tick. Runs immediately **before** `control`, which reads
+/// [`Player::follow_forward`] as one more term of its forward axis — the reference's shape exactly,
+/// where follow pushes the same move-forward bit the W key does.
+///
+/// The player's own turn input runs *after* this in the same frame and therefore wins; that is what
+/// makes the turn-away cancel reachable at all.
+pub(super) fn steer_follow(
+    time: Res<Time>,
+    mut follow: ResMut<FollowState>,
+    mut player: ResMut<Player>,
+    speed: Res<MoveSpeed>,
+    index: Res<GuidIndex>,
+    transforms: Query<&Transform>,
+) {
+    player.follow_forward = false;
+    let Some(guid) = follow.guid else { return };
+    // Re-resolved every tick, nothing cached — a followee that streams out ends the follow
+    // (`0x610e40` → `0x6106e7`).
+    let Some(target) = index
+        .0
+        .get(&guid)
+        .and_then(|e| transforms.get(*e).ok())
+        .map(|t| t.translation)
+    else {
+        info!("follow: the followee is gone — follow ends");
+        follow.stop();
+        return;
+    };
+    let delta = target - player.pos;
+    let flat = Vec3::new(delta.x, 0.0, delta.z);
+    let distance = flat.length();
+    if distance < f32::EPSILON {
+        return;
+    }
+    let bearing = bearing_to(flat);
+    // The INFERRED turn-away cancel (module header): only armed once the facing has come inside the
+    // band, so beginning a follow while facing away turns us around instead of cancelling.
+    let off_bearing = wrap_pi(bearing - player.face_yaw).abs();
+    if off_bearing < CANCEL_ANGLE {
+        follow.aligned = true;
+    } else if follow.aligned {
+        info!("follow: turned away from the followee — follow ends");
+        follow.stop();
+        return;
+    }
+    player.face_yaw = steer(player.face_yaw, bearing, time.delta_secs());
+    let moving = should_move(follow.moving, distance, speed.value);
+    if moving != follow.moving {
+        info!(
+            "follow: {} at {distance:.1} yd (arrive {:.1}, resume {:.1}, speed {:.1})",
+            if moving { "running" } else { "arrived" },
+            arrive_distance(speed.value),
+            resume_distance(speed.value),
+            speed.value,
+        );
+    }
+    // `WOW_FOLLOW_TRACE=1` — the field instrument for "follow won't catch up / overshoots": one
+    // line a second carrying the closing distance and the ground we actually covered, so the
+    // travel rate is a measured number rather than an end-to-end guess (decision 0404: timing and
+    // feel are measured, never eyeballed). Gated because this one IS per-frame.
+    if follow_trace_on() {
+        let now = time.elapsed_secs_f64();
+        if now - follow.traced_at >= 1.0 {
+            // Quiet until a tick has seeded a reference point: an instrument whose opening line is
+            // garbage is worse than one that says nothing for a second.
+            if let Some(from) = follow.traced_pos {
+                let elapsed = now - follow.traced_at;
+                let moved = player.pos.distance(from);
+                info!(
+                    "follow-trace: {distance:.1} yd to go, covered {moved:.1} yd in {elapsed:.2} s \
+                     ({:.1} yd/s), moving={moving}",
+                    moved / elapsed as f32,
+                );
+            }
+            follow.traced_at = now;
+            follow.traced_pos = Some(player.pos);
+        }
+    }
+    follow.moving = moving;
+    player.follow_forward = moving;
+}
+
+/// `WOW_FOLLOW_TRACE=1` — see [`steer_follow`]. One `OnceLock` read per tick when unset.
+fn follow_trace_on() -> bool {
+    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
+    *ON.get_or_init(|| std::env::var_os("WOW_FOLLOW_TRACE").is_some())
+}
+
+/// Register the follow mode. The steer runs in the input stage just before `control`, so the flag
+/// and the facing it writes are what the controller reads this same frame.
+pub(super) fn plugin(app: &mut App) {
+    app.init_resource::<FollowState>()
+        .add_message::<FollowRequest>();
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn the_band_is_three_and_four_point_five_yards_at_normal_speed() {
+        assert_eq!(arrive_distance(7.0), 3.0);
+        assert_eq!(resume_distance(7.0), 4.5);
+    }
+
+    #[test]
+    fn resume_scales_with_speed_but_never_below_the_base_band() {
+        // Above base speed both thresholds stretch…
+        assert_eq!(arrive_distance(14.0), 6.0);
+        assert_eq!(resume_distance(14.0), 9.0);
+        // …below it, arrive shrinks but resume is FLOORED at the 1.0 scale (the `max` in `0x610bfd`),
+        // so the band can never invert.
+        assert!(arrive_distance(3.5) < 3.0);
+        assert_eq!(resume_distance(3.5), 4.5);
+    }
+
+    #[test]
+    fn the_hysteresis_band_latches_on_both_edges() {
+        // Moving: keep going until strictly inside arrive (which is inclusive, so 3.0 stops).
+        assert!(should_move(true, 3.001, 7.0));
+        assert!(!should_move(true, 3.0, 7.0), "arrive is inclusive");
+        // Stopped: nothing happens until resume, so the gap between 3.0 and 4.5 is dead in BOTH
+        // directions — that gap is the whole point of the band.
+        assert!(!should_move(false, 4.0, 7.0));
+        assert!(should_move(false, 4.5, 7.0));
+    }
+
+    #[test]
+    fn bearing_agrees_with_the_controllers_own_forward_vector() {
+        // The invariant that keeps this from drifting out of sign with `control`.
+        for (dx, dz) in [
+            (0.0, -1.0),
+            (1.0, 0.0),
+            (0.0, 1.0),
+            (-1.0, 0.0),
+            (3.0, -4.0),
+        ] {
+            let delta = Vec3::new(dx, 0.0, dz);
+            let yaw = bearing_to(delta);
+            let fwd = Quat::from_rotation_y(yaw) * Vec3::NEG_Z;
+            let want = delta.normalize();
+            assert!(
+                (fwd.x - want.x).abs() < 1e-5 && (fwd.z - want.z).abs() < 1e-5,
+                "yaw {yaw} should point at ({dx}, {dz}), got ({}, {})",
+                fwd.x,
+                fwd.z
+            );
+        }
+    }
+
+    #[test]
+    fn the_turn_is_rate_limited_and_has_a_deadzone() {
+        // A quarter turn at 180°/s takes 0.5 s, so one 0.1 s tick covers 18°, not the whole 90°.
+        let after = steer(0.0, PI / 2.0, 0.1);
+        assert!(
+            (after - TURN_RATE * 0.1).abs() < 1e-6,
+            "clamped to the budget"
+        );
+        // Within the budget it lands exactly on the bearing rather than overshooting.
+        assert!((steer(0.0, 0.05, 1.0) - 0.05).abs() < 1e-6);
+        // Inside the deadzone it does not move at all.
+        assert_eq!(steer(1.0, 1.0 + TURN_DEADZONE / 2.0, 1.0), 1.0);
+    }
+
+    #[test]
+    fn steering_takes_the_short_way_round() {
+        // Bearing just past -π from a facing just under π: the turn must be a small positive step,
+        // not a near-full sweep back the other way.
+        let after = steer(PI - 0.05, -PI + 0.05, 1.0);
+        assert!(
+            wrap_pi(after - (PI - 0.05)) > 0.0,
+            "should wrap forward across ±π"
+        );
+    }
+}
@@ -379,6 +379,12 @@ pub(crate) struct Player {
    /// places the other forward sources do — the direction vector, the wire flags, the swim
    /// amounts, and the turn-rate's "am I translating" test.
    pub(super) autorun: bool,
+    /// **`/follow` is holding the forward key this frame** (decision 0890). Not a mode of its own:
+    /// the reference's follow owns no translation and simply pushes the same move-forward bit the W
+    /// key does (`0x60e790`), so this folds into [`forward_axis`] beside `autorun` and the
+    /// both-button run rather than driving the mover itself. Rewritten every frame by
+    /// [`super::follow::steer_follow`], which runs just before the controller reads it.
+    pub(super) follow_forward: bool,
    /// Free-fly (`F`): the camera moves on its own and the avatar/server position is frozen.
    pub(crate) detached: bool,
    /// Feet position in **Bevy** coords (converted to raw WoW only when sending to the server).
@@ -0,0 +1,533 @@
+//! Selection **by name** — the shared resolver behind `/target` and `/assist` (decision 0886).
+//!
+//! The reference has ONE resolver, `0x493aa0`, parameterised per caller; `/target`, `/assist` and
+//! `/follow` are the same search with different arguments (wow-re
+//! `object-layer/scratch/targeting-by-name.md`, §5-cross-checked):
+//!
+//! | binding | typemask | filter mode | exact-only |
+//! |---|---|---|---|
+//! | `TargetByName 0x489d60` | 8 = UNIT (creatures **and** players) | 0 | Lua arg #2, never supplied |
+//! | `AssistByName 0x489c40` | 0x10 = **PLAYER only** | 0 | hardcoded 0 |
+//! | `FollowByName 0x489ec0` | 0x10 = PLAYER only | 2 (`CanAssist` + alive) | Lua arg #2 |
+//!
+//! **Filter mode 0 is the whole point of the shape**: it underflows `0x493e40`'s jump table
+//! (`dec eax; cmp eax,3; ja`) into the default arm `mov eax,1; ret 4` — *accept unconditionally*. So
+//! by-name selection has **no** range limit, **no** facing cone, **no** dead/reaction/attackability
+//! test, **no** scene-attach visibility gate, and **no** self-exclusion. It is a far wider net than
+//! TAB ([`super::scan`]), which shares none of those defaults. `/target <your own name>` self-targets
+//! in the real client, and so does it here.
+//!
+//! ## The comparison — two tiers, both case-insensitive
+//!
+//! 1. **Whole-string** (`0x64a4c0` → CRT `0x414310`, which folds both operands in its own bytes).
+//! 2. **Longest common PREFIX** (`0x493cb6`-`0x493cf6`), anchored at index 0 — a prefix, never a
+//!    substring. Live whenever the second Lua argument is absent, which is every shipped call site,
+//!    so `/target Rag` really does select "Ragnaros". The running best length starts at **1**, so one
+//!    folded character of overlap is enough to enter the ranking — and a query sharing no first
+//!    letter with anything matches nothing.
+//!
+//! Within one prefix-length class the tie-break is the **strictly nearest** 3D centre-to-centre
+//! dist² (the reference's `fcomp`/`jp` edge is a strict `<`, so ties keep the incumbent).
+//!
+//! ## The one deliberate deviation — and why
+//!
+//! In the reference a tier-1 whole-string hit **returns 0 from the walk callback, terminating the
+//! enumeration** (`0x493ca2`, stopped at `0x4683b7`). Two consequences fall out, both
+//! enumeration-order dependent: among several identically-named creatures the client takes
+//! *whichever the ClntObjMgr table walk reaches first* — never the nearest — and a near "Bobby"
+//! recorded before a far "Bob" beats it outright for the query `Bob`.
+//!
+//! **We do not reproduce the early-out**, and since 0886 was written the order it depends on has
+//! been derived, so the choice is now made against a known reference rather than an unknown one.
+//! Table #1 is an intrusive doubly-linked list (link field at `obj+0x38`, written by the ctor at
+//! `0x4650c2`) whose insert is a **tail append** (`0x4646c3`-`0x4646d4`), so the walk runs
+//! **oldest-linked first** — and "linked", not "spawned": a unit that streams out and back in is
+//! re-appended to the tail. The reference's tier-1 pick is therefore *the same-named candidate that
+//! has been continuously known longest*, which no client-side heuristic reproduces without modelling
+//! the list itself.
+//!
+//! [`Rank::beats`] is instead a **total order** — exact beats any prefix, then longer prefix, then
+//! strictly nearest — which is deterministic and picks the nearest of several same-named kobolds.
+//! wow-re's note calls the reference behaviour here "the single behaviour most likely to look like a
+//! bug in a reimplementation". Reproducing it is now *possible* (stamp each net entity with a
+//! monotonic link sequence, re-stamped on stream-in, and order the walk by it); it is not done
+//! because nearest is the better answer, not because the order is unknown. This is the one place to
+//! change if that judgement is ever reversed.
+//!
+//! ## Not carried
+//!
+//! The reference pre-scans the **party** (`0x493b20`) and **raid** (`0x493b9d`) rosters before the
+//! object walk, matching group members against the guid-keyed name cache at a sentinel distance of
+//! 2.5e9 — so a groupmate with no streamed object is still selectable. benilla cannot hold a
+//! guid-only selection at all yet ([`Selection::target`] is an `Entity`; the same gap already
+//! no-ops `TargetUnit("partyN")` on an out-of-range member, see [`super::click`]), so the pre-scans
+//! wait on that slice rather than being half-built here.
+
+use std::cmp::Ordering;
+
+use bevy::ecs::system::SystemParam;
+use bevy::prelude::*;
+
+use benilla_protocol::guid;
+
+use crate::names::NameCache;
+use crate::net::{Guid, GuidIndex, NetCommands, ObjectStore, SelfPlayer};
+
+use super::{scan, Selection};
+
+/// Which guid families a search accepts — the reference's `edx` typemask argument.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub(crate) enum NameSearch {
+    /// Typemask 8: creatures **and** players (a player object carries the UNIT bit). `/target`.
+    AnyUnit,
+    /// Typemask 0x10: players only. `/assist` — so `/assist <creature name>` cannot resolve, while
+    /// assisting a *targeted* creature can (the no-argument form never goes through the resolver).
+    PlayerOnly,
+}
+
+impl NameSearch {
+    fn accepts(self, guid_val: u64) -> bool {
+        match self {
+            Self::AnyUnit => guid::is_player(guid_val) || guid::is_creature_or_pet(guid_val),
+            Self::PlayerOnly => guid::is_player(guid_val),
+        }
+    }
+}
+
+/// `/target <name>` — resolve and select (`TargetByName`).
+#[derive(bevy::ecs::message::Message, Clone, Debug)]
+pub(crate) struct TargetByNameRequest {
+    pub(crate) name: String,
+}
+
+/// `/assist [name]` — select the basis unit's own target (`AssistByName` / `AssistUnit("target")`).
+/// `None` is the bare form, whose basis is the current selection.
+#[derive(bevy::ecs::message::Message, Clone, Debug)]
+pub(crate) struct AssistRequest {
+    pub(crate) name: Option<String>,
+}
+
+/// How well one candidate answered the query. Ordered by [`Rank::beats`].
+#[derive(Clone, Copy, Debug, PartialEq)]
+struct Rank {
+    /// A tier-1 whole-string (case-insensitive) match.
+    exact: bool,
+    /// Folded common-prefix length, capped by both strings — the reference's
+    /// `matchLen = searchCursor - ctx.name`.
+    prefix: usize,
+    /// 3D centre-to-centre squared distance from the active player.
+    dist2: f32,
+}
+
+impl Rank {
+    /// Strictly better than the incumbent? Exact beats any prefix; then the longer prefix; then the
+    /// strictly nearer. Ties keep the incumbent, and a NaN distance never displaces — both edges the
+    /// reference's `fcomp` + `test ah,5` + `jp` reject makes.
+    fn beats(self, other: Rank) -> bool {
+        match (self.exact, other.exact) {
+            (true, false) => true,
+            (false, true) => false,
+            _ => match self.prefix.cmp(&other.prefix) {
+                Ordering::Greater => true,
+                Ordering::Less => false,
+                Ordering::Equal => self.dist2 < other.dist2,
+            },
+        }
+    }
+}
+
+/// The folded common-prefix length of `query` and `name` — the reference's lockstep walk
+/// (`0x493cc0`-`0x493cee`) through the ASCII case fold `0x41089b`, stopping at whichever string ends
+/// first or at the first folded mismatch.
+fn common_prefix_len(query: &str, name: &str) -> usize {
+    query
+        .bytes()
+        .zip(name.bytes())
+        .take_while(|(q, n)| q.eq_ignore_ascii_case(n))
+        .count()
+}
+
+/// Score one candidate against the query, or `None` when it does not match at all. The reference's
+/// running best length seeds at **1**, so a zero-length overlap is not a match.
+fn rank(query: &str, name: &str, dist2: f32) -> Option<Rank> {
+    if query.eq_ignore_ascii_case(name) {
+        return Some(Rank {
+            exact: true,
+            prefix: query.len(),
+            dist2,
+        });
+    }
+    let prefix = common_prefix_len(query, name);
+    (prefix >= 1).then_some(Rank {
+        exact: false,
+        prefix,
+        dist2,
+    })
+}
+
+/// Everything a by-name search reads. Shared by the `/target` and `/assist` drains.
+#[derive(SystemParam)]
+#[allow(clippy::type_complexity)] // one bundled system param — the app's convention for big query sets
+pub(crate) struct ByNameScan<'w, 's> {
+    /// Every known unit INCLUDING our own avatar — the reference has no self-exclusion on this path.
+    units: Query<
+        'w,
+        's,
+        (
+            Entity,
+            &'static Guid,
+            &'static Transform,
+            Option<&'static ObjectStore>,
+        ),
+    >,
+    self_q: Query<
+        'w,
+        's,
+        (
+            Entity,
+            &'static Guid,
+            &'static Transform,
+            Option<&'static ObjectStore>,
+        ),
+        With<SelfPlayer>,
+    >,
+    /// Read-only: a by-name sweep must never fire a name query per candidate. A unit whose name has
+    /// not been cached yet is simply not matchable — the same limit the reference has, whose own
+    /// name getter reads an already-resolved string.
+    names: Res<'w, NameCache>,
+    /// Only [`Filter::AssistableAlive`] reads these — the reaction ladder behind `CanAssist`.
+    factions: Option<Res<'w, super::Factions>>,
+    reputations: Res<'w, crate::net::Reputations>,
+}
+
+/// The per-candidate filter the reference hands `0x493e40` as its mode argument.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum Filter {
+    /// **Mode 0** — and the shape is the point: `0-1` underflows the `cmp eax,3; ja` bound, so the
+    /// jump table is skipped entirely for `mov eax,1; ret 4`. Accept unconditionally. `/target`,
+    /// `/assist`.
+    AcceptAll,
+    /// **Mode 2** (`0x493eca`) — `CanAssist 0x6066f0` **and** `UNIT_FIELD_HEALTH > 0`. `/follow`
+    /// resolves only to living, assistable candidates, so `/follow <an enemy player>` finds nothing.
+    AssistableAlive,
+}
+
+impl ByNameScan<'_, '_> {
+    /// The active player's position — the distance origin (the reference returns `(0,0)` outright
+    /// when the active player object does not resolve, `0x493ae0`).
+    fn origin(&self) -> Option<Vec3> {
+        self.self_q
+            .single()
+            .ok()
+            .map(|(_, _, tf, _)| tf.translation)
+    }
+
+    /// The reference's mode-2 arm (`0x493eca`): `CanAssist` — the reaction ladder at **≥ 4**
+    /// (friendly; wow-re `w2d1` `CanAssist_Unit 0x6066f0`), the mirror of [`scan::can_attack`]'s
+    /// `≤ 3` — **and** alive.
+    fn assistable_alive(&self, store: Option<&ObjectStore>) -> bool {
+        if store.is_some_and(|s| s.0.unit_is_dead()) {
+            return false;
+        }
+        let self_store = self.self_q.single().ok().and_then(|(_, _, _, s)| s);
+        super::ring_reaction(
+            self.factions.as_deref(),
+            &self.reputations,
+            store,
+            self_store,
+        ) >= 4
+    }
+
+    /// Resolve a name to a unit, or `None`. See the module header for the full law.
+    ///
+    /// Every call logs its verdict under the `by-name:` prefix — one line per typed command, so a
+    /// "`/target` won't take that mob" report is diagnosable from the log rather than re-guessed
+    /// (the `WOW_TAB_TRACE` instrument's shape, ungated because the cost is per-command, not
+    /// per-frame). The line names how many candidates were eligible, which won, and on what.
+    fn resolve(&self, query: &str, search: NameSearch, filter: Filter) -> Option<(Entity, u64)> {
+        let query = query.trim();
+        if query.is_empty() {
+            return None;
+        }
+        let Some(origin) = self.origin() else {
+            info!("by-name: \"{query}\" — no active player object; nothing resolves");
+            return None;
+        };
+        let mut considered = 0usize;
+        let mut nameless = 0usize;
+        let mut best: Option<(Entity, u64, Rank, String)> = None;
+        for (entity, guid, tf, store) in &self.units {
+            if !search.accepts(guid.0) {
+                continue;
+            }
+            if filter == Filter::AssistableAlive && !self.assistable_alive(store) {
+                continue;
+            }
+            let Some(name) = self.names.peek(guid.0) else {
+                nameless += 1;
+                continue;
+            };
+            considered += 1;
+            let dist2 = tf.translation.distance_squared(origin);
+            let Some(r) = rank(query, name, dist2) else {
+                continue;
+            };
+            if best.as_ref().is_none_or(|(_, _, b, _)| r.beats(*b)) {
+                best = Some((entity, guid.0, r, name.to_string()));
+            }
+        }
+        match &best {
+            Some((_, guid, r, name)) => info!(
+                "by-name: \"{query}\" ({search:?}) -> \"{name}\" guid {guid:#x} \
+                 ({}, prefix {}, {:.1} yd) over {considered} named candidates ({nameless} unnamed)",
+                if r.exact { "exact" } else { "prefix" },
+                r.prefix,
+                r.dist2.sqrt(),
+            ),
+            None => info!(
+                "by-name: \"{query}\" ({search:?}) -> NO MATCH over {considered} named candidates \
+                 ({nameless} unnamed); target left untouched"
+            ),
+        }
+        best.map(|(e, g, _, _)| (e, g))
+    }
+}
+
+/// Drain `/target <name>`: resolve, then commit through the shared SetSelection path so a by-name
+/// switch obeys the same stop → select → re-swing law a click does ([`scan::commit`]).
+///
+/// A miss leaves the current target **untouched** — verified: neither failure edge in `0x489db4`
+/// calls `SetSelection`, so `/target nosuchname` never clears what you had. The reference does emit
+/// a game message there (ids `0x127` name-not-found / `0xb8` empty name), but the id→string table is
+/// runtime-populated BSS and wow-re could not statically recover the text, so we say nothing rather
+/// than invent a line. That silence is the known deviation on this path.
+pub(super) fn target_by_name_requests(
+    mut requests: MessageReader<TargetByNameRequest>,
+    scan_params: ByNameScan,
+    mut commit: SelectCommit,
+) {
+    for request in requests.read() {
+        let Some((entity, guid)) =
+            scan_params.resolve(&request.name, NameSearch::AnyUnit, Filter::AcceptAll)
+        else {
+            continue;
+        };
+        commit.commit(entity, guid);
+    }
+}
+
+/// Drain `/assist [name]`: find the **basis** unit, read its `UNIT_FIELD_TARGET` (+0x28 off the unit
+/// block — wow-re proved the offset arithmetic against HEALTH and DYNAMIC_FLAGS on the same line),
+/// and select whatever it is pointing at.
+///
+/// The bare form is `AssistUnit("target")` in the reference — assist whoever you have selected,
+/// creature or player. The named form resolves **players only**. A basis with no target is a
+/// completely silent no-op (verified: the shared tail bails before any send).
+///
+/// **The `assistAttack` leg is correctly absent.** With that CVar (`0xb4d8f8`) non-zero the
+/// reference's tail also opens the swing on the newly-selected target (`0x5ecb70` →
+/// `CMSG_ATTACKSWING`) — but its **registered default is `"0"`** (VERIFIED at the registration
+/// bytes: `0x48fc50`, default-value slot `[ebp+0xc]` = `0x82e570` = `"0"`, int-parsed into
+/// `[cvar+0x28]` by `0x63e127`). So stock `/assist` selects and does **not** swing, which is exactly
+/// what this does. The leg becomes reachable only if benilla ever grows the CVar itself.
+pub(super) fn assist_requests(
+    mut requests: MessageReader<AssistRequest>,
+    scan_params: ByNameScan,
+    index: Res<GuidIndex>,
+    mut commit: SelectCommit,
+) {
+    for request in requests.read() {
+        // The basis: a named PLAYER, or — bare — whatever is currently selected.
+        let basis = match &request.name {
+            Some(name) => scan_params
+                .resolve(name, NameSearch::PlayerOnly, Filter::AcceptAll)
+                .map(|(e, _)| e),
+            None => commit.selection.target,
+        };
+        let Some(basis) = basis else {
+            info!("by-name: /assist — no basis unit; nothing to assist");
+            continue;
+        };
+        // Its own target. A zero/absent guid is the silent no-op — the reference's shared tail
+        // bails here before any send.
+        let Some(guid) = commit
+            .stores
+            .get(basis)
+            .ok()
+            .and_then(|s| s.0.unit_target())
+            .filter(|g| *g != 0)
+        else {
+            info!("by-name: /assist — the basis unit is targeting nothing; silent no-op");
+            continue;
+        };
+        // "Select if it resolves" (`0x489a40`): only a streamed unit can be selected — the
+        // reference's group-roster fallback for an unstreamed guid needs the guid-only selection
+        // benilla does not have yet (see the module header).
+        let Some(entity) = index.0.get(&guid).copied() else {
+            info!(
+                "by-name: /assist — the basis is targeting guid {guid:#x}, which is not streamed"
+            );
+            continue;
+        };
+        info!("by-name: /assist -> the basis unit's target, guid {guid:#x}");
+        commit.commit(entity, guid);
+    }
+}
+
+/// Drain `/follow [name]`: resolve the subject and hand it to [`crate::player`], which owns the
+/// motion (decision 0890). Nothing goes on the wire — follow is client-side movement only.
+///
+/// The two forms differ in more than the argument. A **name** goes through the resolver with the
+/// reference's filter mode 2 — players only, and only ones that are **alive and assistable** — so
+/// `/follow <an enemy>` finds nothing. The **bare** form is `FollowUnit("target")`, a unit-token
+/// resolve that never touches the by-name path at all, so it follows whatever you have selected,
+/// creature or player, with no such filter.
+pub(super) fn follow_requests(
+    mut requests: MessageReader<crate::player::FollowRequest>,
+    scan_params: ByNameScan,
+    selection: Res<Selection>,
+    mut follow: ResMut<crate::player::FollowState>,
+) {
+    for request in requests.read() {
+        let resolved = match &request.name {
+            Some(name) => scan_params
+                .resolve(name, NameSearch::PlayerOnly, Filter::AssistableAlive)
+                .map(|(_, guid)| guid),
+            None => selection.guid,
+        };
+        match resolved {
+            Some(guid) => {
+                info!("follow: now following guid {guid:#x}");
+                follow.start(guid);
+            }
+            None => {
+                // The reference's `if GetSlashCmdTarget(msg)` guard and the resolver's own miss both
+                // land here: nothing to follow, and whatever we were following is left alone.
+                info!("follow: nothing to follow");
+            }
+        }
+    }
+}
+
+/// The shared commit tail — everything [`scan::commit`] needs, bundled, so both drains route a
+/// resolved unit through the one SetSelection path (dedup + the stop → select → re-swing law).
+#[derive(SystemParam)]
+#[allow(clippy::type_complexity)] // one bundled system param — the app's convention for big query sets
+pub(crate) struct SelectCommit<'w, 's> {
+    selection: ResMut<'w, Selection>,
+    net: Res<'w, NetCommands>,
+    me: Query<
+        'w,
+        's,
+        (
+            &'static Guid,
+            Option<&'static ObjectStore>,
+            Has<crate::creature_anim::Engaged>,
+        ),
+        With<SelfPlayer>,
+    >,
+    stores: Query<'w, 's, &'static ObjectStore>,
+    factions: Option<Res<'w, super::Factions>>,
+    reputations: Res<'w, crate::net::Reputations>,
+}
+
+impl SelectCommit<'_, '_> {
+    fn commit(&mut self, entity: Entity, guid: u64) {
+        let me = self.me.single().ok();
+        // The new-target classification `scan::commit` expects from its callers (it does not
+        // re-derive one) — the same `can_attack` the cursor and the TAB scan pass.
+        let attackable = scan::can_attack(
+            self.stores.get(entity).ok(),
+            self.factions.as_deref(),
+            &self.reputations,
+            me.and_then(|(_, store, _)| store),
+        );
+        scan::commit(
+            &mut self.selection,
+            &self.net,
+            entity,
+            guid,
+            me.is_some_and(|(_, _, engaged)| engaged),
+            me.map(|(g, _, _)| g.0),
+            attackable,
+        );
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn exact(dist2: f32) -> Rank {
+        Rank {
+            exact: true,
+            prefix: 3,
+            dist2,
+        }
+    }
+    fn prefix(prefix: usize, dist2: f32) -> Rank {
+        Rank {
+            exact: false,
+            prefix,
+            dist2,
+        }
+    }
+
+    #[test]
+    fn prefix_length_folds_case_and_stops_at_the_shorter_string() {
+        assert_eq!(common_prefix_len("rag", "Ragnaros"), 3);
+        assert_eq!(common_prefix_len("RAG", "ragnaros"), 3);
+        assert_eq!(
+            common_prefix_len("Ragnarosx", "Ragnaros"),
+            8,
+            "capped by the candidate"
+        );
+        assert_eq!(
+            common_prefix_len("zzz", "Ragnaros"),
+            0,
+            "no shared first letter"
+        );
+    }
+
+    #[test]
+    fn a_match_needs_one_folded_character() {
+        // The reference seeds its best length at 1, so zero overlap is not a match at all.
+        assert!(rank("zzz", "Ragnaros", 1.0).is_none());
+        assert_eq!(rank("rag", "Ragnaros", 1.0).map(|r| r.prefix), Some(3));
+    }
+
+    #[test]
+    fn whole_string_match_is_case_insensitive_and_ranks_exact() {
+        let r = rank("kobold vermin", "Kobold Vermin", 9.0).expect("matches");
+        assert!(r.exact, "tier 1 is case-insensitive whole-string");
+        // A multi-word creature name is the case that made the whole-argument trim load-bearing.
+    }
+
+    #[test]
+    fn exact_beats_any_prefix_and_nearest_breaks_the_tie() {
+        // Our deliberate deviation: the reference's enumeration-order early-out would let a nearer
+        // "Bobby" beat a farther "Bob"; we rank exact first, unconditionally.
+        assert!(exact(100.0).beats(prefix(3, 1.0)));
+        assert!(!prefix(3, 1.0).beats(exact(100.0)));
+        // Among exact matches — several identically-named kobolds — the nearest wins.
+        assert!(exact(4.0).beats(exact(9.0)));
+        assert!(!exact(9.0).beats(exact(4.0)));
+    }
+
+    #[test]
+    fn longer_prefix_beats_nearer_shorter_one() {
+        // The reference's `jg -> accept` edge: a longer prefix wins outright, distance irrelevant.
+        assert!(prefix(5, 900.0).beats(prefix(3, 1.0)));
+        assert!(!prefix(3, 1.0).beats(prefix(5, 900.0)));
+        // Equal prefix length falls through to strictly-nearest.
+        assert!(prefix(3, 4.0).beats(prefix(3, 9.0)));
+    }
+
+    #[test]
+    fn ties_keep_the_incumbent_and_nan_never_displaces() {
+        // The reference's `jp` reject edge makes the accept a strict `<`.
+        assert!(!prefix(3, 4.0).beats(prefix(3, 4.0)));
+        assert!(!exact(4.0).beats(exact(4.0)));
+        assert!(!prefix(3, f32::NAN).beats(prefix(3, 4.0)));
+    }
+}
@@ -44,6 +44,7 @@ use crate::interact::{InspectMode, WorldClick, WorldRightClick};
 use crate::net::{ClientCommand, Guid, NetCommands, ObjectStore, SelfPlayer};
 use crate::schedule::WorldStage;

+mod by_name;
 mod click;
 mod cursor_mode;
 mod flash;
@@ -63,6 +64,9 @@ pub(crate) use flash::CombatFlash;
 // The action layer's "attack pressed with no target" request — the nearest-enemy auto-acquire
 // (`scan`) answers it with the same core TAB uses.
 pub(crate) use scan::{can_attack, AttackNearestRequest};
+// The chat layer's by-name selection asks (`/target`, `/assist` — decision 0886), answered by the
+// shared resolver the reference parameterises per caller.
+pub(crate) use by_name::{AssistRequest, TargetByNameRequest};
 // The byte-verified reaction decode + its faction catalog, reused by the unit-frame feed to tint the
 // target's name plate (`TargetFrame_CheckFaction`) the same way the selection ring colours itself.
 // `duel_rung` is the diagnostic face of the same walk, for `/reaction` (decision 0637).
@@ -171,6 +175,8 @@ impl Plugin for TargetPlugin {
            .init_resource::<CombatFlash>()
            .init_resource::<scan::TabHistory>()
            .add_message::<AttackNearestRequest>()
+            .add_message::<TargetByNameRequest>()
+            .add_message::<AssistRequest>()
            .add_systems(
                Startup,
                (
@@ -211,6 +217,19 @@ impl Plugin for TargetPlugin {
                    click::act_on_right_click,
                    click::clear_target_requests,
                    click::target_unit_requests,
+                    // The chat layer's by-name asks (decision 0886), beside the UI's token asks:
+                    // both are non-mouse selection writers committing through `scan::commit`.
+                    // The chat layer's by-name asks, grouped as one chained element (the outer
+                    // chain is at Bevy's 20-tuple limit): `/target` and `/assist` commit through
+                    // `scan::commit` like every other non-mouse selection writer, and `/follow`
+                    // resolves its subject for `crate::player` to move (decision 0890) without
+                    // touching `Selection` at all.
+                    (
+                        by_name::target_by_name_requests,
+                        by_name::assist_requests,
+                        by_name::follow_requests,
+                    )
+                        .chain(),
                    scan::auto_acquire_attacker,
                    scan::tab_target,
                    scan::acquire_and_attack,
@@ -71,6 +71,9 @@ pub(super) enum SlashIndex {
    LootFfa,
    LootRoundRobin,
    LootMaster,
+    Target,
+    Assist,
+    Follow,
 }

 impl SlashIndex {
@@ -105,13 +108,16 @@ impl SlashIndex {
            Self::LootFfa => "LOOT_FFA",
            Self::LootRoundRobin => "LOOT_ROUNDROBIN",
            Self::LootMaster => "LOOT_MASTER",
+            Self::Target => "TARGET",
+            Self::Assist => "ASSIST",
+            Self::Follow => "FOLLOW",
        }
    }

    /// Every registered index — the registry proper. A reference command NOT in this list resolves
    /// nowhere and answers `HELP_TEXT_SIMPLE`, exactly as an unknown command does in the reference
    /// (better than a registered handler that silently does nothing).
-    const ALL: [Self; 28] = [
+    const ALL: [Self; 31] = [
        Self::Reply,
        Self::Join,
        Self::Leave,
@@ -140,6 +146,9 @@ impl SlashIndex {
        Self::LootFfa,
        Self::LootRoundRobin,
        Self::LootMaster,
+        Self::Target,
+        Self::Assist,
+        Self::Follow,
    ];
 }

@@ -122,15 +122,23 @@ pub(super) struct ChatProbes<'w, 's> {
    guids: Res<'w, crate::net::GuidIndex>,
 }

-/// The two setters `DoEmote` drives besides the packet, bundled (the drain is at Bevy's 16-param
-/// ceiling): the **posture** (`EmoteSpecProc == 1` → `SetStandState`) and the **stow**
-/// (`SetSheatheState(0, SNAP)`, unconditional on every emote that passes the gates — wow-re
-/// `sheath-policy.md` §1, site `0x5ef630`). Both are queues into their subsystem's one setter,
-/// never applied here.
+/// Everything the drain **queues into another subsystem's one setter** rather than applying itself,
+/// bundled (the drain is at Bevy's 16-param ceiling).
+///
+/// - `stand`/`sheath` — the two setters `DoEmote` drives besides the packet: the **posture**
+///   (`EmoteSpecProc == 1` → `SetStandState`) and the **stow** (`SetSheatheState(0, SNAP)`,
+///   unconditional on every emote that passes the gates — wow-re `sheath-policy.md` §1, site
+///   `0x5ef630`).
+/// - `target`/`assist` — the by-name selection asks (decision 0886), answered by
+///   [`crate::target`]'s shared resolver so they commit through the same SetSelection path a click
+///   does. Chat never writes [`Selection`] itself.
 #[derive(bevy::ecs::system::SystemParam)]
-pub(super) struct EmoteOut<'w> {
+pub(super) struct ChatOut<'w> {
    stand: MessageWriter<'w, crate::player::StandStateRequest>,
    sheath: MessageWriter<'w, crate::creature_anim::SheathRequest>,
+    target: MessageWriter<'w, crate::target::TargetByNameRequest>,
+    assist: MessageWriter<'w, crate::target::AssistRequest>,
+    follow: MessageWriter<'w, crate::player::FollowRequest>,
 }

 // One parameter per concern — the chat drain fans out to every command's consumer.
@@ -156,7 +164,7 @@ pub(super) fn drain_chat_input(
    mut go_targets: MessageWriter<crate::creature_anim::SpellGoTargets>,
    // The command table (decision 0881) — the reference's own aliases, resolved at boot.
    table: Res<super::commands::SlashCommands>,
-    mut emote_out: EmoteOut,
+    mut chat_out: ChatOut,
    probes: ChatProbes,
 ) {
    let ChatProbes {
@@ -495,6 +503,36 @@ pub(super) fn drain_chat_input(
            ParsedChat::Forfeit => {
                script.queue_duel_request(benilla_ui::script::DuelRequest::Cancel);
            }
+            // The by-name selection pair (decision 0886). Both hand the name to `crate::target`'s
+            // shared resolver — the reference's own `0x493aa0`, parameterised per caller — so the
+            // commit goes through the one SetSelection path a click, TAB and `TargetUnit` share.
+            //
+            // A bare `/target` reproduces `GetSlashCmdTarget`'s fallback (your current target's
+            // name, iff it is a player) and is therefore a no-op re-select; a bare `/target` with a
+            // creature selected resolves to nothing, exactly as the reference's `if
+            // GetSlashCmdTarget(msg)` guard does.
+            ParsedChat::Target { name } => {
+                if let Some(name) =
+                    name.or_else(|| target_player_name(&selection, &mut names, &commands))
+                {
+                    chat_out
+                        .target
+                        .write(crate::target::TargetByNameRequest { name });
+                }
+            }
+            // A bare `/assist` is the ref's `AssistUnit("target")` — assist whoever is selected,
+            // creature or player — so it passes `None` straight through rather than taking the
+            // player-name fallback. The ref reaches the same unit by resolving its NAME first
+            // (`AssistByName`), which can pick a different same-named player; going by the live
+            // selection is the same intent without that ambiguity (recorded in 0886).
+            ParsedChat::Assist { name } => {
+                chat_out.assist.write(crate::target::AssistRequest { name });
+            }
+            // `/follow` (decision 0890) — the subject is resolved by `crate::target` and the motion
+            // is `crate::player`'s; chat only carries the ask. Bare follows the current selection.
+            ParsedChat::Follow { name } => {
+                chat_out.follow.write(crate::player::FollowRequest { name });
+            }
            // The ref's handler is `SlashCmdList["PVP"] = function() TogglePVP() end` — one line
            // over the same binding the popup row calls, so it enters the same intent queue
            // (decision 0646 §3; the `/duel` reasoning above, verbatim).
@@ -565,7 +603,7 @@ pub(super) fn drain_chat_input(
                // sheathes in the reference too. The anim layer's one setter owns the idempotency
                // refusal, so an already-stowed body costs nothing.
                if let Ok((entity, _, _)) = self_player.single() {
-                    emote_out.sheath.write(crate::creature_anim::SheathRequest {
+                    chat_out.sheath.write(crate::creature_anim::SheathRequest {
                        entity,
                        state: 0,
                        ceremony: false,
@@ -577,7 +615,7 @@ pub(super) fn drain_chat_input(
                // the one setter in `crate::player`, which sends `CMSG_STANDSTATECHANGE`, holds the
                // local commit until the echo lands, and runs the sit-stow rider.
                if let Some(state) = posture {
-                    emote_out
+                    chat_out
                        .stand
                        .write(crate::player::StandStateRequest { state: state as u8 });
                }
@@ -159,16 +159,42 @@ pub(in crate::ui_chat) enum ParsedChat {
    /// kick, leave, loot settings, raid marks — apply to the local mirror, so the whole menu
    /// surface is exercisable serverless.
    PartyTest { arg: String },
+    /// `/target [name]` — select by name (`TargetByName`, decision 0886). Resolves creatures AND
+    /// players, case-insensitively, by whole name or longest common prefix, with no range, cone,
+    /// liveness or hostility gate at all. `None` = bare, which the ref's `GetSlashCmdTarget` turns
+    /// into your current target's name when that target is a player (a no-op re-select), and into
+    /// nothing otherwise.
+    Target { name: Option<String> },
+    /// `/assist [name]` — select whatever the basis unit is targeting. A named basis resolves
+    /// **players only** (typemask 0x10); bare assists your current target, creature or player.
+    Assist { name: Option<String> },
+    /// `/follow [name]` — auto-follow (decision 0890). A named subject resolves players only, and
+    /// only living assistable ones (the ref's filter mode 2); bare follows the current selection,
+    /// creature or player. Client-side movement — nothing goes on the wire.
+    Follow { name: Option<String> },
    /// A slash line matching neither a chat command nor an `EmotesText` name — dropped.
    Unknown,
 }

-/// The name half of the ref's `GetSlashCmdTarget` (ChatFrame.lua:650-658): a non-empty argument
-/// is the name (first word — 1.12 names carry no spaces); empty defers to the dispatch's
-/// player-target fallback.
+/// The name half of the ref's `GetSlashCmdTarget` (ChatFrame.lua:650-658) — the argument grammar
+/// **thirteen** shipped commands share, `/target` and `/assist` among them.
+///
+/// Its first line is a trim, not a tokenizer: `gsub(msg, "(%s*)(.*[^%s]+)(%s*)", "%2", 1)` keeps
+/// **everything between the outer whitespace**, internal spaces included. This used to take the
+/// first word instead, on the reasoning that 1.12 player names carry no spaces — true, but
+/// `/target` resolves *creature* names too, and "Kobold Vermin" is the common case. `args` reaches
+/// us already trimmed ([`parse_line`]), which is exactly what that gsub computes.
+///
+/// Empty defers to the caller's fallback (a bare party command acts on the selected PLAYER —
+/// `target_player_name` in [`super`], the ref's `UnitIsPlayer("target")` branch).
+///
+/// **Not carried: the unit-token expansion.** The ref's tail expands `player`, `target`,
+/// `^party[1-4]` and `^raid[0-9]` through `UnitName` before returning, so `/assist party1` means
+/// "that member's name". benilla feeds only the `player` and `target` tokens to the VM's unit
+/// store, so the roster half has nothing to resolve against yet; the tokens would pass through as
+/// literal names. Named in decision 0886 as a gap rather than half-built here.
 fn slash_target_name(args: &str) -> Option<String> {
-    let name = args.split_whitespace().next().unwrap_or("");
-    (!name.is_empty()).then(|| name.to_string())
+    (!args.is_empty()).then(|| args.to_string())
 }

 /// Parse a trimmed slash line into an ACTION (`/join`, `/afk`, `/random`, `/wave`, the dev
@@ -271,6 +297,17 @@ fn slash_command(index: SlashIndex, args: &str) -> ParsedChat {
        S::Duel => ParsedChat::Duel {
            name: slash_target_name(args),
        },
+        // The by-name selection pair (decision 0886). Both take the whole trimmed argument —
+        // `/target Kobold Vermin` is one name, not a name plus junk.
+        S::Target => ParsedChat::Target {
+            name: slash_target_name(args),
+        },
+        S::Assist => ParsedChat::Assist {
+            name: slash_target_name(args),
+        },
+        S::Follow => ParsedChat::Follow {
+            name: slash_target_name(args),
+        },
        S::DuelCancel => ParsedChat::Forfeit,
        S::Pvp => ParsedChat::Pvp,
        // The social verbs (decision 0668). The argument is passed WHOLE: `/who`'s is a filter
@@ -602,20 +602,60 @@ fn real_alias_table_resolves_the_shipped_commands() {
            "{line} resolves to an emote"
        );
    }
-    // …and the four names the reference has NO command for, which the DBC-name match used to
-    // accept. `/follow` is the sharp one: it fired a text emote where the real client's
-    // `SlashCmdList["FOLLOW"]` follows your target (benilla has no follow yet, so it is honestly
-    // unknown rather than wrongly an emote).
-    for line in ["/follow", "/joke", "/puzzle", "/attackmytarget"] {
+    // …and the three names the reference has NO command for, which the DBC-name match used to
+    // accept as emotes.
+    for line in ["/joke", "/puzzle", "/attackmytarget"] {
        assert_eq!(parse_line(line), ParsedChat::Unknown, "{line}");
    }
+    // `/follow` was the sharp one in that class: the DBC-name match fired a text emote where the
+    // real client's `SlashCmdList["FOLLOW"]` follows your target. 0881 made it honestly unknown;
+    // 0890 makes it the real command, over all three shipped aliases (SLASH_FOLLOW1-6 → `/f`,
+    // `/follow`, `/fol`).
+    for line in ["/follow", "/f", "/fol"] {
+        assert_eq!(
+            parse_line(line),
+            ParsedChat::Follow { name: None },
+            "{line}"
+        );
+    }
+    assert_eq!(
+        parse_line("/follow Probeone"),
+        ParsedChat::Follow {
+            name: Some("Probeone".into())
+        }
+    );
    // A command whose handler benilla does not register answers like any unknown command.
    assert_eq!(parse_line("/ginvite"), ParsedChat::Unknown);
+    // The by-name selection pair (decision 0886) — every shipped alias, and the whole-argument
+    // grammar that makes a multi-word creature name ONE name. `/tar` and `/a` are the short forms
+    // the shipped strings carry (SLASH_TARGET2/4, SLASH_ASSIST1/3).
+    assert_eq!(
+        parse_line("/target Kobold Vermin"),
+        ParsedChat::Target {
+            name: Some("Kobold Vermin".into())
+        },
+        "the argument is trimmed WHOLE — `GetSlashCmdTarget`'s gsub, not a first-word split"
+    );
+    assert_eq!(
+        parse_line("/tar   Hogger  "),
+        ParsedChat::Target {
+            name: Some("Hogger".into())
+        }
+    );
+    assert_eq!(parse_line("/target"), ParsedChat::Target { name: None });
+    assert_eq!(
+        parse_line("/a Bob"),
+        ParsedChat::Assist {
+            name: Some("Bob".into())
+        }
+    );
+    assert_eq!(parse_line("/assist"), ParsedChat::Assist { name: None });
    // The whole shipped surface, so a table that half-loaded fails loudly: **225 distinct emote
    // commands** over the 169 `EmotesText` names (the strings repeat — `EMOTE87_CMD1` and `_CMD2`
-    // are both "/sit" — and EMOTE27 "UNUSED" has no row, so it contributes none), and **55 distinct
-    // aliases** across the 28 registered `SlashCmdList` indices.
-    assert_eq!(table.counts(), (55, 225), "(slash, emote) aliases");
+    // are both "/sit" — and EMOTE27 "UNUSED" has no row, so it contributes none), and **62 distinct
+    // aliases** across the 31 registered `SlashCmdList` indices (0886 added TARGET's `/target`
+    // `/tar` and ASSIST's `/assist` `/a` to 0881's 55; 0890 added FOLLOW's `/f` `/follow` `/fol`).
+    assert_eq!(table.counts(), (62, 225), "(slash, emote) aliases");
 }

 // ── The send-side posture-eligibility gate (`emote_send_eligible`) — the director-verified rows
@@ -209,6 +209,40 @@ pub(super) fn drain_container_uses(
            }
            continue;
        }
+        // The openable fork: a clam, an unlocked lockbox, a wrapped gift. `CMSG_OPEN_ITEM`, not
+        // `CMSG_USE_ITEM` — the server's `HandleOpenItemOpcode` is the only handler that answers
+        // with `SendLoot(item guid, LOOT_CORPSE)`, i.e. a loot window over a thing in your bag;
+        // sending USE_ITEM instead casts the item's (absent) on-use spell and nothing happens,
+        // which is exactly the "no way to open clams" symptom. The predicate is
+        // `ItemInfo::openable` — literally the same call the tooltip's `<Right Click to Open>`
+        // line makes, so the promise and the packet cannot drift apart.
+        //
+        // It sits ABOVE the readable fork because the reference's own tooltip law makes the two
+        // mutually exclusive with **openable winning** (wow-re `tooltip-content-law.md`
+        // §1-OPENABLE, the `jmp` past the READABLE test): a wrapped gift that also carries letter
+        // text opens, it does not read. Below the merchant/bank intercepts, which outrank every
+        // ordinary click.
+        let openable = self_q
+            .iter()
+            .next()
+            .and_then(|store| slot_guid(&store.0, bag, slot0.unwrap_or(0), &items))
+            .and_then(|guid| {
+                let obj = items.object(guid)?;
+                let inst_flags = obj.item_flags().unwrap_or(0);
+                let entry = obj.object_entry()?;
+                items
+                    .template(entry, guid, &commands)?
+                    .openable(inst_flags)
+                    .then_some(guid)
+            });
+        if let Some(guid) = openable {
+            debug!("ui_items: open item {guid:#x} (lua bag {bag} → wire {bag_index}/{wire_slot})");
+            let _ = commands.0.send(ClientCommand::OpenItem {
+                bag_index,
+                slot: wire_slot,
+            });
+            continue;
+        }
        // The readable fork: an item INSTANCE carrying `ITEM_FIELD_ITEM_TEXT_ID` (a mail-made
        // permanent letter) opens the reader instead of a use — client-side, no permission
        // packet (vmangos' `CMSG_READ_ITEM` handler gates on the *template*'s PageText, which is
@@ -53,3 +53,94 @@ fn the_whole_shipped_manifest_loads_without_errors() {
    s.resolve();
    assert!(s.errors().is_empty(), "script errors: {:?}", s.errors());
 }
+
+/// **The `$parentTextureFrame` idiom's contract, over the whole shipped UI**: a frame whose art is
+/// meant to cap a unit frame must sit at a strictly HIGHER frame level than that frame's status
+/// bars — because frame level is the only key term above the draw layer, and the layer would
+/// otherwise lift an ARTWORK bar fill straight over BACKGROUND art.
+///
+/// Decision 0884 pinned the layer as bucket-wide and above the frame. Two files had been getting
+/// this right for the wrong reason — `UnitFrames.xml` and `PartyFrame.xml` both *declared* the
+/// TextureFrame after the bars and leaned on the retired key's insertion-order tie-break — so both
+/// inverted the instant the key was corrected, and both reached the director's screen. The
+/// reference spends a real frame level on this in both of its own spellings (`TargetFrame.lua`
+/// l.32-34's explicit `SetFrameLevel(textureFrame-1)`, `PlayerFrame.xml` l.50-52's two anonymous
+/// nesting frames); benilla now uses the first, everywhere.
+///
+/// Name-driven on purpose, so a *new* copy of the idiom is covered the day it is written rather
+/// than when someone remembers to test it. A geometric sweep was run once over this same UI —
+/// every substantially-overlapping cross-frame quad pair at one `(strata, level)`, i.e. every pair
+/// the draw layer alone orders — and found no third instance; the surviving 45 pairs were all
+/// benign (action-bar and bag-slot chrome adjacency, text over art, and a bar fill over its own
+/// frame's trough, which is the intended order). It is not kept as a gate: frozen, that list would
+/// churn on every action-bar edit, and a noisy gate is a gate nobody reads.
+#[test]
+fn every_texture_frame_outranks_its_status_bars() {
+    use benilla_ui::order::unpack;
+
+    let mut s = benilla_ui::script::UiScript::new().unwrap();
+    s.set_screen_size(1024.0, 768.0);
+    let failures = super::load_default_ui(&s);
+    assert!(failures.is_empty(), "manifest load errors: {failures:#?}");
+    // Every unit-frame family painting at once: the frames hide themselves without a unit, and a
+    // hidden frame emits no quads to read a level from.
+    for unit in ["player", "target", "party1", "party2", "party3", "party4"] {
+        s.set_unit(
+            unit,
+            Some(benilla_ui::script::UnitState {
+                exists: true,
+                name: Some("Someone".into()),
+                health: 60,
+                max_health: 100,
+                level: 60,
+                power_type: 0,
+                power: 60,
+                max_power: 100,
+                ..benilla_ui::script::UnitState::default()
+            }),
+        );
+    }
+    s.fire_event("PLAYER_ENTERING_WORLD", vec![]);
+    s.fire_event("PARTY_MEMBERS_CHANGED", vec![]);
+    s.resolve();
+
+    // owner frame name → its (strata, level), read off the packed draw key the renderer sorts by.
+    let mut level_of: std::collections::BTreeMap<String, (u8, u16)> =
+        std::collections::BTreeMap::new();
+    for q in s.extract() {
+        if let Some(name) = s.quad_owner_name(q.target) {
+            let p = unpack(q.z);
+            level_of.insert(name, (p.strata, p.level));
+        }
+    }
+
+    let mut checked = 0;
+    for (texture_frame, &(tf_strata, tf_level)) in &level_of {
+        let Some(base) = texture_frame.strip_suffix("TextureFrame") else {
+            continue;
+        };
+        for suffix in ["HealthBar", "PowerBar", "ManaBar"] {
+            let bar = format!("{base}{suffix}");
+            let Some(&(bar_strata, bar_level)) = level_of.get(&bar) else {
+                continue;
+            };
+            assert_eq!(
+                bar_strata, tf_strata,
+                "{bar} and {texture_frame} must share a strata for the level to decide"
+            );
+            assert!(
+                tf_level > bar_level,
+                "{texture_frame} (level {tf_level}) must outrank {bar} (level {bar_level}): \
+                 tied, the draw layer lifts the bar's ARTWORK fill over the frame's BACKGROUND art"
+            );
+            checked += 1;
+        }
+    }
+    // Player + target (health/power) and four party members (health/mana) — never let this pass by
+    // matching nothing, which is exactly how a renamed frame would silently retire the check.
+    assert!(
+        checked >= 12,
+        "only {checked} texture-frame/bar pairs checked — the name sweep found nothing"
+    );
+    assert!(s.errors().is_empty(), "script errors: {:?}", s.errors());
+}
@@ -1055,3 +1055,145 @@ fn target_frame_border_follows_the_classification_law() {

    assert!(s.errors().is_empty(), "script errors: {:?}", s.errors());
 }
+
+/// The recessed-bar look, asserted in the draw list: the ring art's metal well-edges must paint
+/// **over** the health/power fills, and the name/level text over both.
+///
+/// This is the shape decision 0884 broke. Our transcription used to get the ring over the bars by
+/// *declaring* `$parentTextureFrame` after the StatusBars, leaning on the old draw key's
+/// insertion-order tie-break between siblings at one `(strata, level)`. 0884 pinned the real law —
+/// the draw **layer** is bucket-wide and outranks the frame — so the bars' ARTWORK fills started
+/// painting over the TextureFrame's BACKGROUND ring art and the frames read as pasted-on slabs.
+/// The fix is the reference's own mechanism (`TargetFrame.lua` l.32-34): push the bars one frame
+/// level *below* the texture frame, where no layer can lift them back over it.
+#[test]
+fn the_ring_art_paints_over_the_bars() {
+    let mut s = UiScript::new().unwrap();
+    s.set_screen_size(1024.0, 768.0);
+    load_unit_frames(&s);
+    s.set_unit(
+        "player",
+        Some(UnitState {
+            exists: true,
+            name: Some("Onemage".into()),
+            health: 60,
+            max_health: 100,
+            level: 60,
+            power_type: 0,
+            power: 60,
+            max_power: 100,
+            ..UnitState::default()
+        }),
+    );
+    s.fire_event("UNIT_HEALTH", vec![ScriptValue::Str("player".into())]);
+    s.resolve();
+    let quads = s.extract();
+
+    // The ring art: `UI-TargetingFrame` exactly — not the `-LevelBackground` / `-Elite` siblings.
+    let ring = quads
+        .iter()
+        .find(|q| {
+            matches!(&q.content, QuadContent::Texture { path: Some(p), .. }
+                    if p.ends_with("UI-TargetingFrame"))
+        })
+        .expect("the player frame's ring art");
+    // The bar fills (both bars share the texture; take the highest z of the two).
+    let fill = quads
+        .iter()
+        .filter(|q| {
+            matches!(&q.content, QuadContent::Texture { path: Some(p), .. }
+                    if p.ends_with("UI-StatusBar"))
+        })
+        .map(|q| q.z)
+        .max()
+        .expect("a bar fill");
+    let name = quads
+        .iter()
+        .find(|q| matches!(&q.content, QuadContent::Text { text: Some(t), .. } if t == "Onemage"))
+        .expect("the name text");
+
+    assert!(
+        ring.z > fill,
+        "the ring art must paint OVER the bar fills (ring z={:#x}, fill z={fill:#x})",
+        ring.z
+    );
+    assert!(
+        name.z > fill,
+        "the name text must paint OVER the bar fills (name z={:#x}, fill z={fill:#x})",
+        name.z
+    );
+    assert!(s.errors().is_empty(), "script errors: {:?}", s.errors());
+}
+
+/// The same recessed-bar law for the PARTY member frames, which carry an independent copy of the
+/// idiom (`PartyFrame.xml`'s `$parentTextureFrame` over `$parentHealthBar`/`$parentManaBar`).
+///
+/// Split from the player/target test on purpose: the two files reach the same look through separate
+/// OnLoads, so one green assertion says nothing about the other. This is the copy 0884 broke in
+/// silence — nobody was in a party when the director reported the player frame.
+#[test]
+fn the_party_art_paints_over_the_bars() {
+    let mut s = UiScript::new().unwrap();
+    s.set_screen_size(1024.0, 768.0);
+    // The loot test's prefix (`loot_tests.rs`): PartyFrame's inline <Script> reads
+    // StaticPopupDialogs, which UiPanels.xml defines, and its per-member dropdown OnLoad walks the
+    // whole popup kit.
+    load_xml(&s, "Fonts.xml");
+    load_xml(&s, "UiPanels.xml");
+    load_xml(&s, "GameTooltip.xml");
+    load_xml(&s, "UIDropDownMenu.xml");
+    load_xml(&s, "UnitPopup.xml");
+    // BENILLA_POWER_COLORS (the shared power-colour table) lives in UnitFrames.xml; the party
+    // frames' Update reads it. The player/target frames it also defines stay hidden here — no unit.
+    load_xml(&s, "UnitFrames.xml");
+    load_xml(&s, "PartyFrame.xml");
+
+    s.set_unit(
+        "party1",
+        Some(UnitState {
+            exists: true,
+            name: Some("Onepriest".into()),
+            health: 60,
+            max_health: 100,
+            level: 60,
+            power_type: 0,
+            power: 60,
+            max_power: 100,
+            ..UnitState::default()
+        }),
+    );
+    s.fire_event("PARTY_MEMBERS_CHANGED", vec![]);
+    s.resolve();
+    let quads = s.extract();
+
+    // Scoped to member frame 1 by owner name, so a stray UI-StatusBar from any other frame in the
+    // manifest can never stand in for the bar under test.
+    let mine = |q: &benilla_ui::script::ExtractedQuad| {
+        s.quad_owner_name(q.target)
+            .is_some_and(|n| n.starts_with("PartyMemberFrame1"))
+    };
+    let art = quads
+        .iter()
+        .find(|q| {
+            mine(q)
+                && matches!(&q.content, QuadContent::Texture { path: Some(p), .. }
+                    if p.ends_with("UI-PartyFrame"))
+        })
+        .expect("the party member frame art");
+    let fill = quads
+        .iter()
+        .filter(|q| {
+            mine(q)
+                && matches!(&q.content, QuadContent::Texture { path: Some(p), .. }
+                    if p.ends_with("UI-StatusBar"))
+        })
+        .map(|q| q.z)
+        .max()
+        .expect("a bar fill");
+    assert!(
+        art.z > fill,
+        "the party art must paint OVER the bar fills (art z={:#x}, fill z={fill:#x})",
+        art.z
+    );
+    assert!(s.errors().is_empty(), "script errors: {:?}", s.errors());
+}