- a reference string is a key, never a Rust literal: display text is a key into `GlobalStrings.lua` in-game or `GlueStrings.lua` on the glue screens, resolved at runtime from the player's own install, and a literal is legitimate only as a fallback beside a lookup — the reference's own `UNKNOWNOBJECT` shape — because a re-typed sentence costs the wording nobody diffed, the surface and the voice a composed string has no catalog record to read off (1770, 1815), and localization; the tripwire in `reference_strings.rs` is the check a comment could never be: it walks the workspace, normalises both sides and fails on any literal matching a shipped value — 273 re-typed literals over 34 files, after calibration threw out three false-positive classes (a path-mangling bug, the 85 fallbacks beside a real lookup, and `#[cfg(test)]` modules whose fake tables match by construction — the first run's 378 would have enshrined ~50 phantom sites); `ALLOWED` is a per-file ratchet whose counts may only go down and whose stale rows fail too, regenerated from a clean tree at 253 over 33 files — the first baseline was captured mid-edit and the land gates failed on a row claiming drift in a file that had none — then 252 over 32 as `ui_trade`'s row retired behind a neighbouring conversion to `ERR_TRADE_BLOCKED_S` (2042); stated in its own doc: it cannot find an INVENTED sentence, which matches nothing; and `benilla_ui::strings::fill` is the one template filler — ordered `%s`/`%d`, positional `%N$s`, `%%` collapsing, a starved specifier copied through visibly so a mis-modelled template looks wrong rather than plausible — replacing eight hand-rolled copies that disagreed, one of them `str::replace` filling every `%s` with the same argument, latent only because no message on that queue carried two yet (2045)
- two families convert, and both assert the identifier, not the sentence: the character-create result line resolves its key off the player's chain with the literal as fallback, and three bugs the 23-arm literal table had hidden come out — `0x36` re-typed without its second sentence ("Please try again during off peak hours."), `0x50` invented (there is no `CHAR_NAME_CONSECUTIVE_SPACES` in 1.12, so it takes the default arm), and `0x4B` naming `CHAR_CREATE_NAME_IN_USE` where the code means `CHAR_NAME_RESERVED` — both read "That name is unavailable" in enUS, so no assertion on displayed text could ever have caught it, and a sentinel test that rewrites `CHAR_NAME_RESERVED` in a copy of the real table proves `0x4B` moves while `0x31` does not; the ratchet's first row retires, 21 → 0; and `SMSG_PARTY_COMMAND_RESULT` returns the catalog message the reference's own jump table names — handler `0x5e68b0`, `dec eax; cmp eax,7; ja <silent>; jmp [4*eax+0x5e6a14]`, results 1–8 onto `ERR_BAD_PLAYER_NAME_S` … `ERR_IGNORING_YOU_S` plus `ERR_INVITE_PLAYER_S` (`0x3a`) and `ERR_LEFT_GROUP_YOU` (`0x42`) on result 0 — so text, surface and sound come off one row: `ERR_PLAYER_WRONG_FACTION` (`0xff`) is the red line where the other nine are chat, and three inputs the reference is DELIBERATELY silent on (result ≥ 9, result 0 outside op {0, 2}, an invite with an empty name) stop printing an invented "Party command failed (N)."; tests assert the message id, since no English distinguishes a right key from a wrong one here, and `ui_party` goes 22 → 12 (2045)
- a hidden `<Model>` tile is parked: every orthographic pane shares the atlas's one camera, deliberately kept active whenever any cell is packed because it is the camera that clears the sheet, so `booth_frozen` — exact for every booth ever built, one camera one scene — answered "drawn" for every pane that had left the paint list, and a hidden autocast-shine pane's four spline emitters kept integrating and pushing quads for the whole 600-frame linger at a cell the packer had since given away; `sync_tiles` gains a sixth stage that parks any live tile with no cell this frame — `AnimParked` on the root, which holds the pose evaluator, the compose, the palette write and the global-sequence bone writes, and `ParticleEmitter::set_frozen` on every emitter — with `scene_frozen` stating the freeze law in one place and the park verdict pinned to the composite by a test, so a repack or a full atlas cannot freeze a visible pane's cloud; the bone global sequences move onto the pane's clock, where the material rows have been since 2019: the kernel's Phase B cursor is `(sceneClock − attachSnapshot) % duration` off the scene that owns the instance (`[model+0x2c]`, the snapshot at `CM2Model+0x68` stored at `0x70eae1`), and a `<Model>` widget owns a private `CM2Scene` at `CSimpleModel+0x314` that only its own advance at `0x76d863` moves — so the ping's spinner holds while hidden and resumes where it stopped, with no accumulator of ours; measured with a cooldown pane holding the atlas alive: twelve hidden shine panes cost 54 effect items / 63 transparent draws before, 34 / 43 after, the pool held rather than drained by design; and the particle census's `gated=` column, which printed `fade.is_some()` and read backwards as the gate verdict, is renamed `has_fade=` (2046)
- a doodad is not an entity: `[def+0xa4]` has exactly four writers image-wide — ctor `0x6a7d73` (0.0), WMO doodad-set commit `0x695bb3` (1.0), ADT MDDF commit `0x6b01bb` (1.0), MCSH refresh `0x698cb4` (0.5) — and the 2.5-producer `0x69e280` and the ramp `0x69e770` are reachable only through the entity-attached `[obj+0xe0]` path, which contradicted the §6 both lit-ground sites cited and a capture with gains up to 2.5, so the premise was marked contested at both sites first (the shader's own note told a future session to cut `min(I, 1)` and `LIT_T` together on its strength); wow-re then settled it — two C++ classes inline the same base and share the offset: `CMapDoodadDef` (vtable `0x810a74`, `"WDOODADDEF"`, size `0x198`, serving ADT MDDF and WMO MODD) and the WENTITY light node (vtable `0x810810`, size `0x108`, off `[obj+0xe0]`) that `0x69e770` ramps at 3.3333/s; on the doodad class `+0xf8`, the field the ramp reads its target from at `0x69e894`, is `m[2][3]` of the world matrix — it has no target field — and six trace frames put 111 MDDF placements over 356 draws at gains only in {0.5, 1.0}, none changing, against an entity walking out of shadow at 0.5 / 0.81 / 1.09 / 1.3367 and rising; so ADT map doodads join the fixed-1.0 family with the exterior WMO props, `ShadeSel::Lit` is entity-only, the retained lane's Matte arm drops its prop requirement, and no pixel changes — the cap was faithful for trees all along and lifting it is now an entity-only look change for the director to call; `WORD_SHADE_LIT` is left as a documented dead bit (2050); and the far water's 1.25 s breath is flattened: the director's recording autocorrelates at lag 75 @ 60 fps = 1.250 s (r = +0.60), the 30-frame ocean flipbook at 24 fps, whose frames do not share a mean — alpha 55.985..57.628, DXT3's 4-bit alpha swinging 1.64 at mip 0, 3.72 at mip 5, 6.38 at mip 6 — so at distance the mipped sheet moves together, 0.8..1.84/255 on screen, fogged out at `farclip` 300 and visible at 777; wow-re proved the reference breathes too (`0x68abab` creates the flipbook with `applyGlobal = 1`, the upload hands the driver `Levels = 9`, LOD bias 0, the loop period `[0x8102b8]` 1.25 s over 30 frames), so ours was faithful and it is flattened anyway by the director's eye: each frame's DC equalised per level and per channel at texture-build time as an exact integer redistribution in ±1 steps on a coprime stride (1.64 / 3.72 / 6.38 / 0.555 → 0.000), water and ocean only since magma and slime take the sheet as their body colour, `WOW_LIQUID_DC=raw` restoring the authored art for the A/B (2051)
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
