- the audio crackle was our own stall sampler: `/usr/bin/sample` suspends every thread of the task for each of its snapshots, CoreAudio's realtime IO thread included, and coreaudiod times the client out — every HAL overload of the session sat inside a sample of our own process (three for three by timestamp), and the watchdog fires on exactly the stalls the crackle lands on (a desktop switch's `nextDrawable` block, a world entry's compile burst); in-frame sampling now waits for a run nobody can hear; and the output stream is ours — kira's renderer runs 256-frame chunks on a realtime render thread (time-constraint policy, joined to the device's IO workgroup) into a lock-free ring, the IOProc copies one device buffer and touches nothing else (≤ 0.01 ms), the device buffer goes 2048 → 512, a lost or changed default device rebuilds the stream (cpal's no-op error callback on macOS had paused it forever), a debug allocator aborts on any allocation inside either realtime path, and per-cycle meters name a skipped cycle's layer; `SoundBufferSize` is registered at the reference's own FMOD dial, and `audiowatch.py` puts the OS side — thread bands, coreaudiod's overload verdicts — on the same timeline
- right-clicking a trainer, questgiver, innkeeper, spirit healer, petitioner, tabard designer or battlemaster sent `CMSG_GOSSIP_HELLO` — eight of the reference's fourteen service arms collapsing into one packet, because the dispatch keyed on the CURSOR kind and Speak covers eight bits; the reference runs two structurally identical ladders over `UNIT_NPC_FLAGS` (one for the cursor at `0x482336`, one for the send at `0x5f0289`), so the ladder is transcribed, first match wins low to high: questgiver → `CMSG_QUESTGIVER_HELLO`, trainer → `CMSG_TRAINER_LIST`, petitioner → `CMSG_PETITION_SHOWLIST`, and the spirit healer and innkeeper send NOTHING on the click — they cache the guid and raise the confirm, whose Lua accept sends with a zero guid — both ghost-gated at entry, so a living player clicking a menuless spirit healer now gets exactly what the reference gives: nothing; the stable-master conjunct and the banker/auctioneer fork fall out of the bit order for free
- the -1 blood id is unfilled, not none, and the tests now name the creatures: -1 splits families that share a directory (QuillBoar beside QuillBoarWarrior, GnollMelee beside GnollCaster) and splits single creatures shipped under two rows (a Baby Murloc is both), so a "bloodless" reading would put non-bleeding quilboar beside bleeding quilboar warriors in one Razorfen room; `violenceLevel` 2 IS the reference's default — formatted out of a per-region cap table at `0x86c3f8`, and the one capped locale is koKR, whose green-blood row the table predicts from bytes alone; a spurt whose victim authors no attachment point is dropped, not relocated, and no creature-keyed suppression exists anywhere on the spurt path (ten stop conditions, none keyed on the creature)
- twenty dead copies of ours (`PanelTemplates_*`, `TabButtonTemplate`, `FauxScrollFrame_*`, `ScrollFrame_OnLoad`, `ScrollingEdit_*`) that a later chain entry had been shadowing are deleted under a gate that fails any such shadow, and deleting them turned 141 tests red on four bugs live since the kit swap: lists populated before showing (the reference gates every scroll-bar write on `IsVisible()`, so both arrows were dead on every faux list), every faux window wired to a `frame.updateFunc` the reference does not have (the bar moved, the list did not follow), a trainer pitch declared `local` where an `<OnVerticalScroll>` chunk cannot see it, and the armed-wheel guard bypassed by the template's own `<OnMouseWheel>`; nine faux windows (Skill, Craft, TradeSkill, Auction, QuestLog, Friends, ScriptLog) are still on `updateFunc` and their lists do not scroll until each is rewired
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
