- the pet action bar is the reference's own `PetActionBarFrame.xml`; our 559-line transcription is deleted, and the reference's bar brings what it had left out — the slide, the Ctrl-held and locked gates on show/hide, the hotkey pass whose string-versus-number comparison leaves every corner empty — with the three pet dialogs moved verbatim into the popup slice; one adapter draws the autocast shine (the reference's `<Model>` draws nothing in this engine, so each button gets the token-texture shine at the reference's own 1.2); the stock file registers four events nothing fired, built rather than listed — `PET_BAR_UPDATE_COOLDOWN` when the ten cooldown triples alone move, `PLAYER_CONTROL_LOST`/`GAINED` on the edge of the control flag (`0x4958e0`'s change-gated fire), `PLAYER_FARSIGHT_FOCUS_CHANGED` on the edge of the farsight descriptor, and `UNIT_FLAGS` on a change of the raw flags (its exact fire law a dispatched question) (1953)
- the skills tab is the reference's own `SkillFrame.xml` at its slot between ReputationFrame and HonorFrame; our 874-line transcription is deleted, the `UNLEARN_SKILL` dialog moves verbatim into the popup slice, and `CancelSkillUps` is built — the page's live Close button calls it before hiding, so every close reaches it; it is the reset leg of the temp-point table (`0x4d3e30`, no arguments, no returns) over a table this client keeps empty by construction, so the reset moves nothing, which is the binding's whole effect on this data; `AddSkillUp`, `RemoveSkillUp` and `BuySkillTier` sit on buttons the reference shows only for skill points or step/rank costs no 1.12 server sends and stay named as open gaps; and the expand tab's fit is the reference's load-time law (`GetTextWidth()+45` once, in OnLoad), so the transcription's zero-measure guard and re-fit are gone (1956)
- a 45-bot Blackrock raid circling with sound on holds 60 fps with the main thread at 6–7 ms while the HUD's ms readout says 17–18: that number is the sum of every thread's CPU per frame (main, render, compute pool, audio), not the frame time — the raid costs +6 ms spread over four threads with no thread owning it; two suspects refuted by measurement: mimalloc as the process allocator (interleaved: no change on any thread, reverted the same hour) and page faults (59 a frame on the probe, tens of microseconds, not the render thread's 2.1 ms of kernel time, which is the Metal driver's own work and scales with draws); so the probe counts draws per phase family (transparent outnumbers opaque, mostly the buff wave's particles and nameplates) and `WOW_PROBE_SHARED_SKIN=1` prices cross-body instancing before it is built: nothing moved, because bevy instances only identical meshes and a geared body's merged geoset set is its own mesh (1954)
- three fixes, each removing work a 15 s sample named: booth cameras skip cluster assignment (no bevy light ever reaches a booth, and the default config rebuilt an empty cluster grid per booth view per frame — 0.2 ms of the raid's main thread), the rig palette's watermark clone reuses the row buffer from two publishes ago instead of calloc'ing the 650 KB table every publish (safe because the GPU only ever receives dirty ranges), and prop probes upload the dirty span rather than every allocated row (circling a city streams props in and out most frames, and each change re-sent hundreds of KB for one row); no bucket of the main thread's profile is above a millisecond — the raid's +2.5 ms there is forty-five bodies' worth of every per-entity lane at ~55 µs a body, and the one lever left that is all ours, one texture array per body folding ~9 draws to ~3, is sized at ~1.5 ms and not started (1954)
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
