Warcraft III renderer loss audit — 8 November 2026

No emulator cause is demonstrated; no emulator repair was made. The retained original establishes an unanswered audio import and disappearance of an OS renderer PID. It does not establish that the disappearing PID owned Warcraft's page. The bounded production audio fixture passes. A remaining prelaunch PID/target mapping gap prevents another original in this task.

Evidence: scratch/runs/20261108-warcraft3-renderer-process-loss/. Read first: released remote journal investigation, its final cause audit, journal tail, process samples, browser logs and driver; also the Warcraft and earlier current-browser notes. Single worker, no nested workers, no game execution or new box in this task.

What the retained original establishes

The compressed journal was streamed through SHA-256 and parsed without a raw expansion. Its 2,582,639,811 original bytes hash to bb37538245ebad990e24f07e91e64c7b43cad35d6e1666cfd0f5458b07639668, matching the release. The 1,950,341-byte controller trace was also independently hashed and inspected. retained-audit.json preserves both hashes and tails.

The final nested chain is WorkerLink._onMessage 2968485 → GuestRpc.invoke.voice_play_ring 2968486 → GuestRpc.host.voice_play_ring 2968487, arguments [720900,24246244,136704,0,2], slot 4, endpoint 5, 04:52:59.323 UTC. There is no retained return/error. The last observer emission at 04:52:59.324 drains through 5,937,009. The released final 218,625-slot contiguous tail remains meaningful; earlier 8,528 lost slots in ten ranges remain a limitation. Older unmatched calls across those ranges are not hangs.

The busiest retained observer epoch second contains 47,620 rows, and the largest batch contains 1,025 rows including a loss marker. Loss means the true earlier peak is not bounded by that measurement. Most entries are unrelated Worker messages; only 17,929 retained host voice_play_ring entries occur in the entire original. Repeating millions of those other records is unnecessary.

PID 11520 is present at 04:52:58.956 and absent at 04:52:59.954, before the 04:53:17.923 owner read and its UNKNOWN. Browser process requests continue. Its dominant renderer CPU time is a correlation. SystemInfo.getProcessInfo has PIDs but no target IDs; page/Worker bindings have target IDs and emulator process ID 1000, but no OS PID. The retained driver captures neither a tracing mapping nor browser target crash/exit events. It handles target detach for Worker bookkeeping without retaining an independent exit record.

Thus exact Warcraft page ↔ PID11520 mapping is unknown, and so is the original renderer exit reason. Original Chrome stderr contains no corresponding fatal stack or exit explanation. The later browser exit code zero describes normal browser closure, not PID11520. OS crashpad/apport/kernel/OOM records were not retrieved by that run; their absence from evidence is not proof that the remote OS had no such records. The archived box was not resumed to seek them. Page and observer emissions both stop, so uninterrupted synchronous execution inside the last audio import is also unproved. Neither audio deadlock, OOM, guest trap, browser crash cause, emulator regression nor historical stall cause follows from this boundary.

Exact production audio path

IDirectSoundBuffer_Unlock in src/09a8-handlers-directx.wat refreshes an already playing looping buffer by calling host_voice_play_ring(handle, wa, size, 0, 2). lib/guest-rpc.js:createWorkerImports uses an asynchronous descriptor refresh only when its live voice map matches both pointer and length. Otherwise it posts a blocking RPC. The final observed RPC means this fast path was unavailable; the retained records do not say why.

GuestWorkerLink._onMessage services the RPC through the main broker, whose invoke calls the selected lib/host-audio.js arrow import. That import calls ctx._voices.playRing(...) and returns zero only afterward. playRing first looks up the voice, pumps notifications, obtains a PCM view and selects strict loop === 2. Refresh then marks audio hot, updates extent, flushes or joins the optional recording tap, and attempts an existing worklet descriptor write. If not already routed, it can promote a live looping BufferSource after lazy module readiness; otherwise it uses the existing delayed-ring or decode/splice fallback. Those branches were not bracketed in the original.

notifyPump walks active notification registrations, computes progress and signals crossed events through the final host namespace. Worklet progress reads try at most four times. route checks routing gates/module readiness, then writes a descriptor or constructs a node. Module loading is asynchronous; this path does not synchronously wait for the module promise. The AudioWorklet uses bounded descriptor retries and finite render quanta. There is no Atomics.wait in these page-side audio files. This source audit does not rule out a native browser failure or establish which refresh branch the original entered. Tap allocation sizes, native calls and event reentrancy remain unobserved in that original; no speculative lock or audio change is justified.

Narrow probes and real-browser fixture

Run-local probes.cjs reuses the released independent observer and tested coalesced recorder. It selects voice lifecycle/notification/worklet methods and unique lexical refresh hot/tap/delay/decode brackets. It excludes unrelated Worker messages. Arguments retain primitive values and object identities; object graphs, getters and thrown-object message getters are not traversed. The diagnostic ring is bounded at 32,768 slots of 2,048 payload bytes, with explicit loss/oversize/fragment markers. The recorder remains bounded at 32 MiB. Probe code adds no guest writes, guest polls, timer or scheduling changes. Existing observer timing and production audio timers remain in use.

probe-test.cjs checks receiver/argument identities, original object and promise results without then assimilation, exact thrown object, exactly-once calls, host zero versus inner undefined, actual production getMemory exception, hostile receiver/error getters, unchanged fixture memory, and fail-closed source anchors. These probes are diagnostic artifacts, not a production source change.

retained-voice-shape.json recovers voice 720900's actual open arguments: 44,100 Hz, stereo, 16-bit. Corrected fixture v4 uses that format, exact pointer 24,246,244, length 136,704, offset zero, voice ID 720900, fixed 512 MiB shared WebAssembly memory and actual browser Worker slot 4. It invokes production GuestRpc.createWorkerImports/createMainBroker, createAudioHost, the worklet router and the actual AudioWorklet processor. No emulated guest program or full interpreter module is substituted into an original run.

Both unobserved and observed arms pass lazy-load fallback, loop=2 promotion, same-node loop=2 refresh, locally read advancing worklet cursor, stop/release, replay/new node and close. Synthetic notifications at zero, half-ring and STOP exercise the active notification/progress path; their registration is not claimed to match the original's unknown notification state. The first promotion travels as blocking RPC; the routed refresh travels as asynchronous call. Outer results remain zero and production lifecycle semantics agree.

Initial iterations with only 100 ms audio startup waits failed a fixture cursor assertion while AudioContext time was also stationary. Retained v1/v2 failures are not emulator failures. A two-second startup allowance followed by a 500 ms cursor interval resolves the fixture assertion. v3 passes at 22,050 Hz; v4 corrects the recovered format and adds active notifications. The initial overbroad probe anchors failed closed and were narrowed before observed runs.

Corrected v4 measurement Result
Synchronous bursts 3 × 900 refreshes
Maximum rows per epoch second 75,925 (> retained original 47,620)
Final contiguous chain 77,682 slots, 38,841 entries + 38,841 returns
Lost / oversized / pending fragments 0 / 0 / 0
Recorder admitted / durable batches 386 / 386
Maximum queued plus in-flight bytes 3,831,102 (< 33,554,432)
Median 900-refresh unobserved / observed burst 6.710 / 170.980 ms
Added time per refresh 182.522 µs

These are sequential synthetic arms, not a randomized original workload: the inner burst takes 25.48× as long with probes. Original-run overhead is unmeasured. Passing beyond the measured rate does not bound every possible original burst. validation.json validates every final sequence increment, parent entry/return pairing and bounded queue. A finite 1.8-second synchronous callback additionally leaves its entry durably readable while still pending; the independent recorder/observer continues and its exact result 37 returns.

Independent exit control and remaining gate

After closing the synthetic audio fixture, a deliberately requested local Page.crash produces Inspector.targetCrashed and Target.targetCrashed(status="crashed", errorCode=134) for exact page target C4BBF068032641CDC37FBA689C88F4BD. Independent browser samples lose renderer PID 3269640; native stderr identifies that same PID and signal 6. Crashpad minidumps for the v3/v4 controls and the v3 apport report were recovered and hash-verified. Those records explain only these deliberately induced local control exits, never the original. A 700 ms observation in v3 was too short; v4's bounded independent samples retain the eventual crash event and PID loss.

Pre-crash Tracing.start/end qualification, attempted with both event delivery and a stream, retained no page marker or frame/PID mapping within its bounded completion wait. Therefore a reliable prelaunch page/Worker-to-OS-PID mapping is still unqualified. Post-crash correlation in a deliberate control does not supply that missing original mapping. Remote burst service and original overhead for this narrower selection are also unmeasured.

This is the concrete remaining gap: qualify that mapping and independently retain browser target exit status plus OS crash/exit diagnostics alongside the inner refresh chain on a subsequent bounded original. The existing original cannot supply missing inner branches or an exit reason retrospectively. No new box was allocated while this prerequisite failed; no replay, WebGL follow-up, guest/config/clock edit, audio disabling or lock removal occurred. Player control, full27, parity, FPS, audio quality and campaign stay open.

Validation, preservation and lifecycle

Focused neighbors pass: audio routing 36/36, Worker audio cursor, DirectSound refresh seam and AudioWorklet processor tests. No emulator source changed, so no failing-before repair claim or new locked-build claim is made. Git references, relevant binary diff, source hashes and owned new-file patches preserve recovery without source/media archives. Closed logs are losslessly compressed and round-trip hash-verified; the predecessor's multi-GB journal was never expanded locally. Run size stays bounded.

All local fixture Chrome processes exit zero on browser closure; profiles and ports are closed. local-shutdown.json records checks. No box was created: box ID null, state NOT_ALLOCATED, expiry null, stop receipt null. No inherited box was touched. Sampled free disk stayed above five billion bytes; the lowest explicit sample was 5,003,784,192 bytes. A continuous disk minimum was not recorded. The coordinator independently restored headroom. No RAM gate, source/media archive, old-box resume/delete, deployment or nested worker.