Warcraft page nonreturn investigation — 8 November 2026
The original synchronous page operation remains unidentified. The new independent journal passes local production controls, but its fresh remote delivery fails before Warcraft startup readiness. This attempt supplies no emulator regression or justified synchronization change. The previous reviewed Prologue loading/cinematic evidence remains the gameplay boundary; Thrall control is unproved.
Evidence: scratch/runs/20261108-browser-page-nonreturn-cause/. The predecessor
capture investigation, its final audit
and trace, and producer-ready gameplay
were read and preserved. No failed command was replayed.
Source audit and diagnostic coverage
The actual synchronous path is GuestWorkerLink._onMessage →
GuestRpc.createMainBroker → its selected host import. serveRpc reads the
parked Worker's numeric arguments, invokes the import and publishes its answer.
The diagnostic brackets both the broker invocation and all five existing host
function call sites, preserving the original undefined receiver. It records
the message parent, slot, import name, exact numeric arguments and object
identities. The original broker still owns dispatch and exception handling.
The VFS path merits observation: lib/filesystem.js's writeGuestBytes copies
through sparse-safe chunks, then notifyGuestWrite calls
ctx.exports.invalidate_code_range(guestAddr >>> 0, size). That exact call is
bracketed, with its original exports receiver and both arguments. VFS
writeFile/setEndOfFile, Worker message callbacks, thread-slot retirement,
workerSyncState, publishWorkerThunkState and host DLL-count publication are
also bracketed. These are coverage choices, not suspected defects.
host.js establishes a distinct host shadow before Worker PE loading and
initializes its metadata with thread ID 7. The current
src/13-exports.wat:3957 invalidation export already saves and enables
d3dim_lazy_bypass, performs page-watch/code invalidation and restores the
prior value. The lazy access/fence paths test that bypass before entering the
shared lazy lock. Neither this audit nor the runtime evidence establishes a
lock cycle or supports removing synchronization. Cache retirement and
page-watch loops were inspected without promoting a speculative repair.
source-callpath-audit.json retains the relevant excerpts and line references.
Run-local HTTP transforms insert synchronous call brackets and append narrow method wrappers. No exported instance or export is replaced. The diagnostic does not write guest memory, poll input, alter guest clocks/scheduling or add page Promise reactions. Wrappers return the original result immediately, including Promise identity; a synchronous return record is not an async completion claim.
A separate observer Worker reads a dedicated 16.8 MB diagnostic
SharedArrayBuffer, never guest memory. Records carry entry/return/error,
wall/monotonic timestamps, receiver identity, arguments and synchronous message
parent. Data descriptors are encoded without invoking getters. Buffer/view
arguments retain exact identity/type/range; their contents are explicitly
unsampled. Large structured messages use bounded fragmentation/reassembly.
Overflow, lost slots and incomplete fragments are explicit, not silent
truncation or "Object". Probe console payloads are strings. The independent
Node recorder writes gzip members and fsyncs each received observer batch.
Local qualification and preserved failures
Before allocation, qualified-software/ and qualified-webgl/ passed through
actual Chrome, production guest/render Workers, the unchanged released
producer/binding helpers and numbered native-input/ACK transport. Both native
red triangle controls were personally reviewed. Strict missing/zero/nonowner
producer, missing context/endpoint, wrong guest, renderer replacement,
postbinding endpoint replacement, stale review, module mismatch and UNKNOWN
controls dispatched zero input. Six positive commands received exact ACKs;
independent native release under delayed DOWN also passed.
The actual observer was attached while paused, enabled and resumed through
CDP. A finite 1.8-second busy callback reached through production
WorkerLink._onMessage left the observer independently readable. Fsynced
samples arrived between entry and return. Observed/unobserved controls preserved
receiver, exactly-once invocation, undefined result and exact thrown object.
Both final arms have zero oversized/lost records and no unfinished fragments.
preallocation-qualification.json records these results.
Failures remain intact: the initial callback expectation missed production's message prefix; the next control referenced an undeclared counter and exited Node, requiring SIGTERM of its own Chrome and verified profile cleanup; an anchor-count assertion expected four host calls instead of the audited five; one software message exceeded the initial record size, requiring fragmentation and complete requalification; a historical-evidence path assertion used the review folder instead of the retrieved folder. These iterations are excluded from qualification. An early remote status read raced directory creation and is also retained.
Local qualification did not qualify the fresh remote transport. No native
trace is used or claimed. Browser SystemInfo.getProcessInfo samples are
streamed before failure; an independent OS sample preserves thread stats,
wait channels and explicit permission-denied kernel-stack reads. These are
process samples, not JavaScript/WASM stack attribution.
Fresh attempt: delivery fails before startup readiness
One UUID-idempotent no-environment box, bx_8yacs6h8, was created at 04:31:05.334 UTC. The supervisor, independent 25-minute watchdog and normal-stop child were observable before allocation. The allocation expiry was 04:56:05.194 UTC; browser stop remained the separately fixed 04:42:43.500 UTC. The unchanged original app/module/media and strict driver were selected for SoftwareGL. There was no native command, screenshot or startup owner binding.
final-cause-audit.json establishes the narrower failure:
- Readiness
Runtime.evaluaterequest 25 was sent at 04:31:53.963 and timed out UNKNOWN at 04:32:14.002, after 20,039 ms. - The page emitted seven heartbeats after that request was sent. The last retained page emission is 04:32:00.238, received at 04:32:53.805. This excludes uninterrupted synchronous page execution throughout that interval; it does not prove entry of the requested body.
- All 1,097 durable observer drains report journal sequence zero, no production entry, no loss/oversize and no pending fragments. Median receipt lag is 32,946 ms, maximum 53,720 ms. Seven browser process samples were recovered; subsequent browser-level requests also missed deadlines.
- The observer drain request after UNKNOWN timed out.
Browser.closewas sent at 04:32:34.014; Chrome actually exited zero at 04:32:34.561, although its ACK timed out at 04:32:54.047. - The guest Worker script was served at 04:31:58.365, but its
Debugger.enableattach handling was not entered until 04:32:46.285, after Chrome had exited. No guest Worker resume/module proof completed. The observer's earlier attach/resume did complete.
The journal/console delivery and Node ingestion path did not retain bounded service on this remote host. The idle observer emits a batch every 10 ms and the receiver synchronously compresses/fsyncs every batch; that is an important unqualified load, not an established slow syscall or Chrome defect. No per-batch service-time measurement separates recorder I/O, WebSocket delivery and inspector event queuing. Consequently this fresh failure cannot be substituted for the historical Warcraft stall or attributed to VFS, invalidation, a lazy lock, renderer or guest code.
The precise remaining gap is remote observer/recorder delivery qualification before original launch, followed by acquisition of an actual production entry without return during the historical loading interval. Idle batch coalescing, a bounded heartbeat, recorder service-time/queue-depth measurements and attach/resume/drain timing need a failing-before/passing-after control before another original. None was silently changed mid-run. First UNKNOWN closed the attempt; no input, replay, second original or WebGL arm followed.
Source, verification and shutdown
Staged Git SHA: 8fdba6283331ae518068cf0260fb9d4ba6e77ca5.
runtime-pins.json, relevant-source.patch, the new boot-file patch and
owned-helper-patches.patch retain exact source references and relevant hunks.
The released binding/producer/observer/input helpers are byte-identical to the
predecessor. All 630 remote runtime pins remained unchanged. Twelve media
files are referenced by registered path/current SHA; four historical hashes
match and eight independent historical hashes remain unavailable. No source
tree or media archive was created.
All 32 remote artifacts / 32 unique hashes were retrieved and independently
reread for size/SHA at 04:33:28.033 UTC, before requesting normal stop.
Image/module/identity/ACK validation passes for the qualified local controls:
eight PNG CRC/IDAT decodes and capture hashes, six exact ACKs and matching
CPU/render module hash
f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121.
The fresh original has no completed module/owner proof and receives no such
claim. Its controller/Chrome PIDs are absent, both ports closed and profile
and writer lock removed.
Normal stop completed at 04:33:30.050 UTC, receipt
stop_69deb02fa52f1fed, forced: false. Final box state is archived,
expiry null. Both lifecycle children exited zero; the supervisor has no
remaining children. Local/remote disk stayed above 5,000,000,000 free bytes.
Own completed logs were losslessly compressed, and identical captured modules
link only to an immutable retrieved module, not the mutable build output.
No emulator code change or build claim, guessed lock removal, RAM gate,
old-box resume/delete, nested worker, public deployment or gameplay progress
claim. Full27, parity, FPS, audio, control and campaign completion remain open.
RECOVERY.md identifies the evidence and prohibits replay/resume.