- black rain is a zero-padded mip: `RainDrop01.blp` stores its 2×16 and 1×8 DXT3 levels as one block each where the block grid needs four and two, our decoder zero-padded the rest — an all-zero BC2 block is black at alpha 0 — and the rain's Mod2x lane reads no alpha, so every far streak that minified onto levels 3 and 4 (three-quarters and half black by the census) multiplied the scene toward black from ~43 yd, where level 3 enters the trilinear blend, to the fog at 70; the reference sizes a stored level at `bpp·max(4,w)·max(4,h)/8` bytes from its file offset and the header's `mipSizes` never sizes an upload, so a short level is completed by the next levels' own grey blocks — `[L3 | L4 | L5 | L6]` — and `Header::dxt_level_span` reads that span on both decode entry points, CPU and native byte-identical, with `pad_to` kept only for a file that ends inside a level, which no shipped asset does (543 of 32,309 DXT files carry a short level, every one non-square, none past EOF); not a mip cap and not a grey fill — the file's next bytes are what the reference draws; wow-re's cross-check settled which arm is which: the block copier `0x5a5780` is live on the shipped Direct3D default (`0x5a0a80` → `0x5a4f60`'s cell `[6][6]`), the upload's `imageSize` at `0x59f4e6`–`0x59f515` over the zero-copy pointer (`0x5a8555`) on OpenGL, the same bytes either way — and the `GL_TEXTURE_MAX_LOD` clamp is dead on every device (its gate `[0xc0f274]` has one writer, a zeroed `ebx`), so every tail level is sampled; why it surfaced now: 1642/1645's isotropic trilinear takes the streak's width for its LOD where the retired anisotropy 8 took its length, which carried the far streaks onto level 3 inside the fog window for the first time; `benilla-extract blp --mips` writes every authored level and prints the per-level texel census, and a real-asset test pins the copy byte for byte (B358/B225, 2020, 2024)
- the cooldown sweep draws again: the extract's `ModelPane` arm drew the atlas cell it found in the bridge, but the cell is packed by `sync_tiles` AFTER the extract and the conversion is memoized on the engine's list — so a cooldown armed on a quiet interface extracted once with no cell, the cell arrived a frame later, nothing re-ran the conversion, and the tile lived on the atlas for the whole cooldown with no quad on screen (the 2013 harness test had re-pinged the minimap to force a second conversion, encoding the defect as the contract); the arm now publishes the request — rect, paint key, alpha, clip beside the unit ladder — and `compose_tiles` appends one premultiplied quad per cell packed THIS frame in the `UiQuadAppend` lane, a function of the bridge alone; the second defect behind it: 2019's rotation rows were uploaded to the world's shared light buffer only, while a tile's material twin binds the tile's own black-light buffer whose mat-anim region is zero-initialised, so every quadrant read the identity — and `cooldown.blp` is 0.6-alpha black outside a four-texel strip, so the sweep would have drawn as a static disc and never cleared; `MatAnimMirrors` now carries the region into every buffer a lane binds under one generation gate; measured six seconds into a 30 s stance cooldown, the button rects read means of 79 and 22 against 150 and 48 before, the top-right wedge cleared from 12 toward 3 o'clock; `WOW_TILE_TRACE=1` is the tile probe (2023)
- a body inside a GameObject's hull is left where the server put it: the reference's resolver is a time-of-impact sweep and nothing else — no push-out stage anywhere in `0x634040`'s closure, and `0x632830`'s clip drops a candidate face whose clipped vertices all sit more than 1/36 yd (`[0x7ff9c8]`) behind the prism's leading plane — where avian's slide opened and closed with a Gauss–Seidel push-out whose minimum translation for a tall capsule in a small box is horizontal, shoving a seated body 0.555 yd off its stool on the first frame after the server seated it (B359); the `depenetrate` stage is deleted rather than disabled, and because the deletion alone left the body unable to walk out (a straddled seat top read as a wall at t = 0) the sweep carries the reference's band as `BACKFACE_BAND`, measured against the face's vertices, so a body walks off a stool and still stands on a sloped floor its feet sit a few centimetres under; the creature clamp cast its down-ray from 2.5 yd above the seat, a probe the reference never casts — its settle sweeps DOWN from the current position (`0x636dcd`–`0x636e45`) and the only way up is the step-up after a blocking horizontal hit — which found a cage's lid as Galen Goodward's floor (B372; seat 22.416, ray from 24.916, lid 24.667): a walker continuing its path now runs the shared `grounded_step` from where it stood with a creature's own 2.0 yd rise budget (`[0x801628]`) and takes only the Y, an idle unit casts down from its seat with the 1/36 band above it, and continuity is one server path long since `0x7c6420` re-bases on every packet; the `clmp` trace tag (`WOW_CLAMP_TRACE`) reads one clamp cast and the `move` tag marks input-less travel as CREEP (2018); and a seated body is not re-faced by the mouse: the reference's mouse-MOVE handler `0x514400` rotates the camera unconditionally and hands its facing to the body only through `0x5145e0`, whose last conjunct is `GetStandState() == 0` (`0x51460c`, the client-predicted cache), with `0x5151b0` re-testing the descriptor byte — so a right-drag while sitting orbits the camera round a body that keeps its facing and sends nothing; benilla restored the facing on the stun/death predicate only and spun a seated body in its chair with every sample, streaming a `SET_FACING` each; `MoverInput::mouse_may_turn_body` carries the stand-state term, and the keyboard turn keeps `may_turn` alone because that path stands the body (2025)
- the vendor window's face and name: the stock `MerchantFrame.lua` takes its title from `UnitName("NPC")` and its face from `SetPortraitTexture(MerchantFramePortrait, "NPC")` — the `"npc"` token, `CGGameUI`'s `[0xb4e2d0]` that every NPC window reads on its own `*_SHOW` — and that seam had four defects stacked on it: `feed_interact_npc` ran a frame ahead of the net apply, so the frame `SMSG_LIST_INVENTORY` raised `MERCHANT_SHOW` still held the previous NPC (the Goldshire inn titled "Innkeeper Farley" over Brog Hamfist's stock); `feed_units`' dirty gate had no interact term, so a swap to a second vendor moved no descriptor and skipped the feed; the memo row was never written, so `UnitExists("npc")` stayed true after a close; and `SetPortraitTexture` stored its token raw where the booth slots are keyed lowercase — the reference resolves every unit token through `0x515970`, whose ten compares are all `_strnicmp`, so the binding folds on the way in; the writer now seats itself `.after(apply_net_updates).before(drive_display_facing)`, pinned in the built schedule graph rather than the source, the gate takes `interact.is_changed()`, the memo is written, and the fold lands at the one seam where a Lua token becomes a slot name — every NPC window reads the same token through the same chain; `WOW_PROBE_VENDOR_SWAP=1` stands between Brog Hamfist and Barkeep Dobbins, 5.45 yd apart, reads the name inside `MERCHANT_SHOW`, opens one over the other and closes it — five FAILs before, five PASSes after (2022)
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
