Comment corrections in crates whose comment pass has already landed. Comment text only: every file strips to identical code.
- Fixes#61. The `<Script file=>` note in `benilla-ui`'s loader now names the install directory as the reference's scan root (`0x646ea0`, from Storm's base path `0xc52418`), not the process's working directory. The deviation it records is unchanged.
- Fixes#96. `liquid.wgsl` no longer labels the ADT water alpha a deviation. The swatch's own alpha is what the reference uses: the `0xc7fbc0` table's texture binds only behind `[0xc800ec]`, which never holds one.
- Fixes#117. `wmo.rs` now gives prop creation as the streaming-window commit (`0x698720` → `0x695aa0`, once per group) and per-frame visibility as the portal walk's admission (`0x685d70` → `0x6838f0`).
- `model.rs`: the billboard-card normal flip carries the Deviation label, with the reason `benilla-formats` already gives where the test lives: the reference shades such a card inverted and camera-dependent.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Part of #41: the comments of `benilla-assets`, Rust and WGSL, rewritten to the rule in `docs/METHOD.md`: what the code does and the reference fact behind it, in a line. 3,872 comment lines become 1,568.
- The citations that were only a private record number go first, mechanically, each edit shown to be a pure removal.
- The rewrite corrects the comments that stated wrong facts: the alpha-key cutoff now cites the reference's per-blend-mode table (`0x85ad20`), the texture-filter note names the subsystems that ignore the global setting, a test doc said an exhausted variation roll returns the last variation (it returns the first), and about twenty more found on the way (renamed or deleted items, a shader flag the doc misread, docs sitting on the wrong function). "Deviation:" now marks only a choice with a reason; an unbuilt behaviour says so plainly.
- Verified: every file strips to identical code, and `cargo fmt --check` is clean. A second reader compared every file old against new, restored the facts the rewrite had dropped, and checked each new citation against the reference.
String literals are untouched; they are #92.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Comments that pointed into the maintainer's private reverse-engineering notes (a note by name, a
path, a section or finding id, the analysis's own name for a client function, a process word) now
carry the fact itself, cited by the reference client's address, by vmangos file and line, or as
"the reference". Nobody outside could follow those pointers.
Comment text only, plus the test-assert and log strings that carried the same pointers. No claim
changes: a comment the notes have since corrected keeps its wording here, and correcting it is
separate work.
The two reference generators, `scripts/gen-reference-globals.py` and
`scripts/gen-reference-verb-events.py`, named the private notes' layout to find their inputs. They
now take the input files as arguments (`--capture`, `--framexml`, `--names`, `--extents`,
`--disasm`). Run on the same inputs, both write the committed tables row for row; only the two
header lines that named the old environment variable change.
Checked per file:
- the code is identical: each file with its comments stripped compares equal before and after; a
changed string literal is compared under a mask that proves only literal text moved and every
literal kept its line count;
- no client address the old text cited was dropped;
- a second reviewer checked every added address against its derivation;
- a detector for the notes' names, paths, ids, function names and process words finds none left;
- `scripts/gates.sh` is green.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- the spell owns its lifecycle state: the in-flight cast, the local cancel, the cooldowns and the spell-modifier tables live in one module (2328)
- a malformed MPQ, DBC, M2 or WDT file is refused instead of crashing the client, nine read sites bounded
- a test that needs the install or the addon corpus fails instead of skipping when the data is there; the Bagnon render suite runs again (2329)
- the dev crash injector is armed on every platform, and no instrument writes a path relative to the working directory
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- a size-less `<Model>` pane's rect is its file's box in layout units: the reference's geometry overrides (`0x76d080`/`0x76d0d0`) answer the bbox extent whenever no size is authored, `bboxExtent · 768·√(a²+1) · layoutScale` FrameXML units (`1280 · extent` at 4:3), written into the pane's layout input when the file's facts land, re-derived on an aspect change, and ended for good by any authored size; `SetPosition` is the same unit, which is how `0x4a7b20` centres the arrow at `½·GetWidth, ½·GetHeight` through the override — so the map arrow's sprite arm and its hand-sized 33.6 px footprint go (its rect is `0.0262 × 0.0263` layout units, 33.5 × 33.7 FrameXML units at 4:3, 41.0 × 41.2 at 16:9), and the stock `ItemAnim` pane plays the item push unadapted (`ReplaceIconTexture` on the type-14 batch, the 1000 ms clamp, the stock `OnAnimFinished`), `ContainerFrameAdapters.xml` losing 0887's Lua card and its hand-measured 42.41 × 124.72 rect — the implicit one at 16:9 — and keeping the keyring fork and the three bag verbs (2015); and the cooldown is the reference's own model: the texture-transform law `uv' = R_q((uv + t − p) ⊙ s) + p`, `p = (½, ½)`, the raw quaternion component-lerped and never normalised, is one function (`uv_transform`) with the shader's UV fold as its twin — rotation and scale tracks bake per file sequence slot beside the translation, the mat-anim table gains an affine row `[cos − 1, sin, sx − 1, sy − 1]` on `anim_slots.z` whose identity is the zero row so every static batch reads exactly what it did, and a UI tile clones the twin material per animated batch and writes both rows every frame off the pane's play head, never the world clock; the stock `Cooldown.xml` runs off the chain on 2007's two handlers, the file's four quadrant quads turned by their own rotation tracks (`0 → −90°` over each quadrant's 250 ms window, clockwise from 12 o'clock) and its alpha tracks revealing the star for the flash; retired: the native `Cooldown` widget and `SetCooldown` (an Era API 1.12 never had — `CreateFrame("Cooldown")` is the reference's unknown-type refusal now), the loader's `Model → Cooldown` substitution, the app's hand-built pie and star, and our `Cooldown.xml`; `assets/ui` is eight files (2019)
- `UpdateMapHighlight` gets its zone-level exits (B360): at zone zoom ours returned the nil tail for every cursor position, so hovering Deadwind Pass from Duskwood showed no label where 1.12 shows "Deadwind Pass"; the reference's `0x4a7fa0` runs two lookups at zone level and answers a NAME only — first the current zone's exploration-gated overlay list (the builder at `0x4a6ad9` gating on the explored byte), each row's HitRect scaled by 1/1002 and 1/668 against the normalized cursor, edges inclusive, the first hit naming its Area1, then the continent grid re-windowed through the displayed zone's own rect, naming the neighbouring zone or city and nothing for the displayed zone itself — and only the name-only tail is reachable from there, fileName nil and the six coords zero, so the stock frame hides the highlight quad, as the director's own 1.12 shot shows; and `SetVerticalScroll` stores the offset as given (B370, B363): the reference's `0x786db0` compares the new offset against the old and writes it to `[+0x328]` verbatim, `0x787100` re-anchors the scroll child off it, and the range at `[+0x320]` is never read on that path — every clamp the reference exhibits is FrameXML's, through the scroll bar's `[min, max]`; ours clamped into `[0, GetVerticalScrollRange()]`, the child's overflow past the FRAME, which is 40 px short of where `FauxScrollFrame_Update` sends the bar on the 220-tall skills list (three rows lost) and 15 px short on the 287-tall who list (one); the setter fires `OnVerticalScroll` only on a change, as the bytes do, `BenillaScroll_Step` bounds its wheel in Lua where the reference's handler does, and the ScrollFrame kind's prose stops calling its offset unpinned (2017)
- overdraw's cheap half, the sky first: since 0588 every sky element wrote `frag_depth = 0.0`, which costs a pipeline its early-Z, so on an outdoor frame every full-screen dome fragment under the ground and the buildings was shaded in full (the gradient's `asin`, `atan2` and eight table lookups at dawn and dusk) and thrown away; one shared `sky_vertex.wgsl` — bevy's mesh vertex verbatim plus `out.position.z = 0.0` — is now the vertex stage of the gradient dome, the cloud dome, the star patches and the celestial discs and glares, and the model lane's `WOW_SKY_DEPTH` branch moves from its fragment to its vertex, so no sky fragment declares `frag_depth` and the rasterizer rejects the covered screen before the shader runs; every sky `specialize` zeroes the rasterizer bias constant the `sky_order` rung had been leaking into `DepthBiasState::constant`, harmless only while the fragment overwrote its depth, and the prose beside the pin follows it; then the retained `static_gx` pass draws BEFORE bevy's opaque pass (`StartMainPass → static_gx → MainOpaquePass`) — buildings and doodads are the frame's best occluders and a terrain fragment its dearest, six samples a pixel — so walls reject the terrain behind them instead of the reverse, the pass taking the first-use clear with it and the dev depth probe ordered after it; measured on 16 capture scenarios: 14 bit-identical, `northshire-dusk` 1 px and `water-night` 11 px, attributed to exact depth ties on static/terrain silhouettes and last-bit interpolation noise; a depth prepass and near-to-far bins are declined until a journal says where the time goes (2016)
- the grass cut goes per pixel because of a quarter-mip: `0x6813d0` sets `D3DSAMP_MIPMAPLODBIAS = +0.25` on stage 0 (`0x6813f4`) once per frame before the detail-doodad pass — capture-confirmed on all 29 doodad batches, and outside the `[0x6b1000, 0x6b3200)` scan 2004 had called complete; it decides the whole look because the `clutter_cut` probe finds 6 of 31 shipped detail atlases carry a BINARY alpha pyramid (mip 0 has DXT3's 16 levels, every level below exactly two, both starting zones among them), so `a × ramp` degenerates to a global on/off and every texel of a leaf crosses at the same depth — Elwynn's survivor coverage is flat at 47.6 % across the whole 52.5–61.1 yd band, then 0.0 %, while the probe's box-filtered control erodes 47.7 → 20.9 over the same band, so the art has the gradient the shipped pyramid threw away, and `clutter_size` puts the median tuft at 0.92 yd tall, 1.03 yd deep, a full yard for the boundary plane to sweep; the escape is bilinear: full alpha needs a 2×2 all-opaque texel neighbourhood, and those run out between mip 2 and 3 (179/56/10/0/0/0 on the dominant Elwynn cell), so the reference's λ ≈ 2.9–3.3 sits past the cliff and the tuft erodes continuously — 2.2 % of a tuft's pixels leave in the last half-yard with the bias, 11.6 % without, 24.2 % one mip sharper; wgpu has no sampler LOD bias, so the clutter lane re-samples base colour at `view.mip_bias + 0.25`, keeping the sampler every other model batch shares at the reference's 0; ruled out on the way: our BLP decoder (all nine uploaded levels byte-identical against the archive's 52 793 BLP2s), mip regeneration, the `has_mipmaps` byte's value 2 (a plain boolean, an authoring marker), anisotropy and geometry; what it does not fix is resolution-dependent, and so is the reference's own look — λ with the bias is 3.22 at the capture's 648p but 2.48 at 1080p and 1.48 at 2160p, so at a modern resolution the reference itself pops leaf by leaf, and scaling the bias with the viewport is one reversible line left for a look call (2021)
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
- the grass fade is a plane across the view, not a ring around the player: `0x6b2b80` texgens the camera-space position through `T(0,0,−52.5)·Ry(π/2)·S(1/17.5)`, so the fade coordinate is view-space depth, `u = (z_eye − 52.5)/17.5` — we measured `distance(eye, fragment)`, a sphere, which meets the ground as a circle centred on the player that slides as they walk and bites a piece out of every tuft it crosses, and that was the hard border; the ramp is `0x6b2320`'s quantised 64×8 table, `alpha = 4·(63 − col)`, read bilinearly as `(254 − 256·u)/255` and capped at `252/255` — plateau to 52.64 yd, zero at 69.86 yd, near grass 98.8 % opaque and never 100 — where we had a clean 1 → 0 line; the product `tex0.a × ramp` is what the 128/255 `detailDoodadAlpha` cutout tests (both stages MODULATE, `[0xc7b2fc]`, confirmed verbatim in all four apitrace captures), so the tuft eroding out at 61.11 yd is the reference's own behaviour and the "fix" that would have removed it never landed; the pass owns its render state — blend mode 2, ALPHAREF 128, cull off, depth-write on, the per-slot draw `0x6b2d60` binding nothing but the texture — where we passed the batch's authored blend and cull through, and the `clutter_state` probe counted the cost: of 322 detail batches, 2 author Opaque (no cutout, their margin drawn as a solid card) and 5 single-sided; the build gate measures the chunk's box against the frustum corner ray derived from the live projection (≈1.31× at 16:9, ≈1.45× at 21:9), not its centre plus a 24 yd margin that popped whole chunks in on relief (2004)
- a `<Model>` widget's animation state is Lua-observable and lives in the engine: a private scene clock its own `OnUpdate` advances by `trunc(elapsed·1000)` ms on visible frames only (`0x76d7f0`), an arm with an anchor the sampler re-reads every frame (`0x7121a0`), `OnUpdateModel` at the top of every paint (`0x76d1a0`) and `OnAnimFinished` from the completion callback (`0x76cdc0`); the file's sequences and bounds are facts the host hands over, `SetModel` seeds Stand now or when they land (`0x70ebd0`) with a queued arm replaying after the seed, and the XML `scale=` on a model pane is the model's scale (`CSimpleModel::LoadXML` `0x76cac0`), which the loader had never read (2007); and the pane renders its M2 — every visible pane with a file gets a cell of one render-target atlas at its device size through one orthographic camera, the root at `1280·modelScale·layoutScale` FrameXML units per model unit, the tile's `AnimationPlayer` paused and seeked to the engine's play head, the light black, particles in their own unit (`768·√(a²+1)` per model unit, neither scale), composited as one premultiplied quad at the callback rank; two corrections off the bytes: `OnUpdateModel` fires only with a model installed (`[widget+0x318] ≠ 0`) and `OnAnimFinished` fires once per arm for a looping sequence too (`0x719370`); the loader now applies a `<Model>`'s `file=` at all, which every XML-declared pane had lacked; retired: the minimap ping's byte-measured sprite, since `MiniMapPing` and `WorldMapPing` draw their own file on the pane's clock (2013)
- the stance feed fires `UPDATE_SHAPESHIFT_FORMS` only for a list change: the reference fires it at `0x4b28ff`/`0x4b2e43` from learn, unlearn and rank-replace alone, while ours fired on any change of the pushed view, and every fire runs the stock `ShapeshiftBar_Update`, which shows `ShapeshiftBarMiddle` unconditionally past two forms and `Show()`s an already-shown frame — no `OnShow`, no manage pass — so a stance switch or the category-47 cooldown arming and expiring left the middle strip drawn alone over a raised bottom-left bar; castability now rides `SPELL_UPDATE_USABLE` (`0x4b31c0`), active/texture/cooldown are pushed silently and repainted on `PLAYER_AURAS_CHANGED` and `SPELL_UPDATE_COOLDOWN`, the feed ordered before both handler sets so they read the fresh list (2009); and the loot-roll frame waits for its item template (B371): `GroupLootFrame_OnShow` paints once and nothing repaints it, and we fired `START_LOOT_ROLL` on the frame the packet arrived, against a snapshot whose name, texture and quality were still `None` — the reference's `0x61b310` ends in the item-template cache lookup `0x55ba30`, a hit firing the event there and a miss firing it from the arrival callback `0x61b460`, so the feed holds a fresh roll's event until the template answers, with two releases that are ours: a negative answer (a frame that never opens cannot even be passed on) and a 2 s deadline paid out of the roll's minute; `arg2` stays the wire's countdown, as `0x61b430` pushes it (2010)
- the FPS journal ships to players: it had registered inside the dev-only `DevToolsPlugin` and been armed by an environment variable, so nothing we shipped could have written a row on the one GPU-bound frame in the corpus, a Steam Deck's at 94 % busy; `/console fpsJournal 1` — a CVar of our own, persisted like every CVar — opens `benilla-config/Diagnostics/fps-journal.csv` and appends one row a second, `0` closes it mid-run, and `WOW_FPS_JOURNAL=<path>` stays as the harness lever; eight GPU columns appended, `gpu_ms` then `gpu_opaque, gpu_static, gpu_transp, gpu_glow, gpu_post, gpu_ui, gpu_other`, read off bevy's `RenderDiagnosticsPlugin`, which times passes where the device offers `TIMESTAMP_QUERY_INSIDE_PASSES` (Vulkan and DX12) and leaves the cells empty on an Apple GPU, empty being the honest value; `static_gx`, the four `ffx_glow_*` passes and `ui_gamma_decode` open diagnostic spans so a city's own draw is named in the split; a fresh file opens with the adapter, backend, driver and `gpu_spans yes|no` line; and a probe's `/console` line, which skips the stock `ChatEdit_ParseText` and had landed in a parser arm that knew only `reloadUI`, now forwards to `ConsoleExec` and writes a CVar the way the typed line does (2008)
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
- right-clicking a vendor with an item on the cursor sells it: `0x5df5d0` asks `GetCursorItem` before anything else and a non-zero answer takes the arm to `CMSG_SELL_ITEM` instead of `CMSG_LIST_INVENTORY`, where benilla sent the list unconditionally — the whole stack, to the clicked vendor, with no merchant-window gate anywhere in the arm; the clear is `SetCursorItem(0, …)` rather than `ClearCursor`, which fires `CURSOR_UPDATE` and leaves the source slot greyed until the server's inventory update (1914); and the NPC you are talking to is not clickable: `0x5f0130` refuses a right-click on the NPC whose window is already open, against a global with exactly two writers — the interaction setter, whose fourteen callers are every one a window opener, and its clear — so the first click always gets through and no double-click is eaten, where benilla re-sent the arm's opcode on every click (1866)
- a Hated or Hostile questgiver's `!` now disappears, every `!` blinks off and back on when you resurrect, and the minimap no longer dots a corpse, your own pet or an untrackable unit: the sweep tears down on HOSTILITY, not on the flag — `0x5eb134` is a convergent block reached from two branches, so three incompatible readings had each seen one path — and an escort giver's `!` clearing with its flag is the field watch instead, which B257 now rests on; the dots gain `0x4eaa90`'s three preconditions (1906); the quest log's rebuild is all-or-nothing — the reference makes every rebuild a no-op until its template misses are answered, so a cold log shows empty and then lands in one step, never our `"..."` row per in-flight quest — sorted level then title, case-insensitively, with the `id == 0` group first under the reference's own "Missing header! (quest designers)"; and `GetQuestLogTitle`'s `isHeader`/`isCollapsed` answer 1/nil, not booleans
- an animation event has a place, not just a time: the M2 event kernel (`0x719370`) places each fired key through its bone, and benilla carried the when and the what and never the where, so every animation-event sound played at the model root — 149 of 244 `$DSL` records sit off their model's origin, out to 67.6 yd on `Maraudon_Waterfall01.m2` — and the point belongs to the record, not the tag, since 75 models author one tag twice; doodad, GameObject and shake sounds now sound where the key is (1904); and the unit dispatcher's arms do not agree about the point they are handed — footsteps and emotes at the event point (the footstep's material foley stays at the origin), `$TRD` at the origin, `$CSD` at attachment 17 rather than its head-mounted records, `$CSS` at the event point on a hit and attachment 1 on a miss — with `0x623b90`'s attachment-else-bias resolver transcribed table and all (1915)
- `GetSkillLineInfo` out of range answers the reference's thirteen-value tuple, numbers in the rank slots, not a single nil — the stock skill window adds two of those slots unguarded straight off `SetSelectedSkill(0)`, so one nil raises the moment it drives us; one shape serves every out-of-range case because the bounds test at `0x4d3675` is a single unsigned compare, and our source had asserted the opposite (1919); the outbound wire is traced by name (`out`, mirroring `in`), a line pushed only after a successful write — a 5-second visit to the Goldshire poster reads as 87 lines (1901); the binding shape gate seats a player and checks the `Unit*` verbs it had been skipping, coverage 220 → 249, all agreeing with the reference (1911); and a UI capture scenario opts the UI in itself instead of photographing a UI-less world and exiting 0, the opt-in answering `false` in a player build that has no harness
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
- the paper doll joins the collapsed rig lane and its body panes render every other frame while open, the pose still evaluating at full rate (boothHalfRate CVar) — the open character window's cost at Goldshire ~+2.7 → ~+1.0 ms
- the parked frame's idle diet: avian's dynamics pipeline goes (the world has zero dynamic bodies), the hidden dev overlay stops drawing an empty pass, and a unit's swim state is re-checked only when it moved — 12.8–13.3 → 12.0 ms parked at Goldshire
- the UI tick stops sweeping the whole frame arena three times over: OnUpdate frames and the message and cooldown kinds live on registries kept at their one write site — idle tick 530 → ~195 µs a frame
- the WOW_MEGA_STATIC bracket retires, past its recorded condition now that the retained pass carries its own
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- every run now merges the static world: never-fade doodads and interior props bake into per-cell blobs, −0.42 ms at the Stormwind gates and ~−0.2 at Goldshire; WOW_STATIC_MERGE=0 opts out
- the fading doodads stay individual — one translucent draw per cell depth-killed whatever feathered behind it, the lamppost popping in behind a Goldshire fence — and WOW_MERGE_FADERS=1 keeps the dev lane
- merged blobs sort at their own centre instead of the world origin, and their pipeline variants pre-warm so no blob compiles live mid-walk and steps the fade
- roaming converges on the parked shape: a quiet cell re-bakes its fragments into one blob, and a doodad straddling two tiles hands off to a loaded neighbour when its owner tile unloads
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- behind WOW_STATIC_MERGE=1, never-fade doodads, fading doodads and interior props bake into one mesh per 133-yard cell and material — −0.99 ms at the Stormwind gates, a wash in Lower Blackrock Spire, GPU flat
- the merged faders fade in-shader on the reference's own fading draw state, so the ramp is as smooth as an individual doodad's
- interior props carry their light-probe slot per vertex, and the arrival cover waits for every unbaked blob so a reveal never shows a hole
- a merge census (WOW_MERGE_CENSUS=1) prices each population before it is built — WMO group geometry merges nothing and stays individual
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- water and lava keep time in the shader: the 24 Hz CPU tick that re-uploaded every liquid material is gone, −0.28 ms at the Stormwind gates, and lava scrolls smoothly instead of stepping
- static transform tracking is pinned on — the adaptive check spent two full scans a frame to say yes in a world where movers are 1–3% of the rows
- WOW_MEGA_STATIC merges static geometry per material and 133-yard cell — an experiment lever, priced at −1.15 ms CPU for +0.3 ms GPU at the Stormwind gates
- WOW_GPU_MS times the whole frame on the GPU and reads 2.2 ms against 17.4 CPU at the Stormwind gates: the frame is CPU-bound, so the work ahead is on the CPU side
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Blackrock Mountain's lava bubbles animate at last: their sprite's flip-book channel was baked from a sequence they never play — UV and tint loops now bake per sequence, and a placement whose sequences disagree gets its own material
- a bonfire's logs were vanishing while its flame burned on: a body planted just under its own floor now claims the room above it, and the flame card and fire sound follow the body's cull instead of outliving it
- a WMO prop's animation host stops being cull-tested as an object in its own right
- WOW_TILE_RADIUS's doc now states its real default, 2, and why
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- a player build ships beside the dev one — dev chords and probe surfaces go dark, and a gate enforces what a player can still reach, by medium not by module
- the binary carries its own shaders and its shipped UI, finds the WoW folder by itself, and keeps its local state in one benilla-config folder
- the glue screens get a capture shutter, and the engine stops depending on the protocol crate
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- lava renders outside Blackrock at last: the mesher was dropping every open-world lava and slime surface, and 28 river mouths get their missing sea back in the same sweep
- Undercity's broken corridor returns, geometry and color both: a one-byte file overrun dropped the whole room, and a paint job one record short was being thrown away entire
- collision goes one-sided like the real client: walls stop you from the front and let you pass from behind
- the Booty Bay entrance is wood and rope again instead of a black cutout, glass picks up a real reflection, and models far from the map's center stop trembling
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- the GCD system now matches the real client end to end: opening a chest no longer eats the running global cooldown, and a running sweep's start stops wobbling between frames
- the mounted left shoulder's wild swing is fixed at the bone: a bone flagged to ignore its parent's rotation now does, decided before any camera-facing runs
- a gameobject effect with several animation sequences plays the right one: Molten Core's rune flame spreads and dances instead of collapsing into a bright dot
- the targeting cursor asks the thing under it and greys by default like the real client, and the ground-target reticle keeps its own location rules
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- the loading screen now compiles every shader the world can reach: sky, water, spell shards and model variants stop hitching the first time they appear
- a locked chest's cursor carries the LOCKED word like the real client, and the lock verdict never leaks into the sent packet
- enchanting asks before it acts: a bind-on-enchant item gets its confirm popup, and so does replacing an enchant that's already there
- a long gossip menu scrolls like the reference with rows grown to their wrapped text, and music already playing is not restarted when the server re-sends it
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- a stealthed unit's particles and ribbons now dim with the body, and a depth-prime pass stops overlapping translucent parts from darkening each other
- changing gear re-dresses the character in place instead of tearing the whole model down and rebuilding it
- terrain tiles furnish cell-by-cell and model render forms build paced across frames, so first contact with fresh ground no longer hitches
- Ghost Wolf renders properly ghostly, and a live toggle swaps its action-bar icon to the active form while it holds
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- stealth reads as stealth now: the body fades to the reference's 30 % and creeps with the proper stealth walk and stand poses
- the item and enchant glow subsystem exists: permanent weapon glows and enchant visuals show in the world and at character select
- the character-select background's colour banding (the reported "16-bit" look) is fixed at the render target, and the tauren stage's black corners with it
- aura body tints got a real per-instance channel, so a spell can tint the whole unit without touching its neighbours
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- terrain draws per chunk now, and one authority culls everything outside: buildings, billboard cards, particle emitters
- the fps and memory creep on long flights is fixed twice over: material animations only tick what something draws, and far-away art is released by distance (with a 30 s dwell floor so fast cameras don't thrash it)
- the second crash ever reported (a cast racing a despawning unit at flight speed) is fixed, with a regression test that reproduces the exact panic
- every build now prints which commit it is, so a bug report can say exactly what it ran
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- An honest correction first: the previous snapshot's message announced the
owned skinning lane, but the code itself missed that snapshot by minutes —
two sessions landed at once and the export carried the older tree. The
skinning lane actually ships here, together with its follow-up.
- The follow-up is the bone-entity collapse. A creature's ~75-bone skeleton
existed as ~75 world entities, each paying transform propagation and
bookkeeping every frame, to serve the ~5 bones anything actually hangs off
(held weapons, glow emitters, the mount seat, overhead markers). A rig's pose
now lives in plain arrays on the model itself; only the consumer anchor
points remain as entities. At the crowded dungeon spot ~59,000 entities
became ~28,000 anchors (80k total → 50k), palette uploads coalesce into a
handful of writes, and another ~3 CPU ms came off the frame.
- The movement echo the server sends about your own character is applied now,
the way the original client always applied it — groundwork that GM flight
cashes in later today.
- The capture harness pins its clock per frame now, not just its frame count —
a slow run used to drift every animation out from under a golden image.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- The dungeon framerate reports were right about the mechanism: the cost is
CPU-side and it scales with NPC count, not with what's drawn. Two per-unit
lanes were each 30–100x more expensive than the work they do. Measured at the
reported Lower Blackrock Spire spot: 29.7 ms/frame, 72 CPU ms, before.
- "Which room is this NPC in" scanned whole dungeon groups triangle by
triangle — a spire stacks several 15k-face groups into any vertical column,
so 37 moving NPCs cost 1.2 million triangle tests per frame. Each group's
faces now sit in a column index built at load; same verdicts (pinned by
exactness tests), 88k tests, the lane drops from 11.3 ms to 1.1 ms.
- Animating one creature cost ~190 microseconds per frame in generic per-bone
animation machinery — hash lookups and boxed curves per bone, mostly to
produce constants. A direct pose evaluator now does the same math into the
same joints for ~16 microseconds per rig, checked bone-for-bone against the
old output across crossfades, overlays and masks. The reference animates
every in-range unit with no view cull, and now so can we: a raid's worth of
animating rigs (~450) costs ~4 CPU ms where it used to cost ~85.
- Net at the reported spot: a flat 60 fps. The performance probes also learned
to report CPU work (ms and %) instead of wall-clock fps, which is what makes
these numbers repeatable on a busy machine — and comparable to the Mac
report's own units.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
