MW3 original endpoint successor — 2026-11-07

The original demo qualified its actual guest Worker and owning presentation, then reached a personally reviewed Zipper Interactive splash. The attempt ended at an operator command-client error before any native input or main-menu navigation. No cockpit, movement, turn, fire, FPS or parity claim is supported.

Evidence: scratch/runs/20261107-mw3-endpoint-original/. The prerequisite mw3-websocket-staging was published and released at a3823fdfec3c59b73bcec2f16be71a04f13c5029, also this run's source SHA. Its documented prepare-original.cjs command passed without pin drift. All 304 individually provisioned files passed remote SHA/length verification before launch and again after shutdown. Generated signatures, API table, vendor closure, genuine MSVCP50/Mech3Msg, original media and normal config were retained. The original route was /index.html?debug&candidates&no-log&gl-renderer=webgl&app=mw3. No generated startup proof or accepted fixture matrix was rerun.

The pinned boot.py constructed the immutable interval 14:39:51.267Z–15:04:51.267Z and launched exactly one original runtime. The repaired waiter serialized a complete valid endpoint. readiness.json and renderer-readiness.json prove actual guest and owning renderer readiness; both recovered live module identities are e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467, 2,182,493 bytes. The renderer reports WebGL 2 / ANGLE Vulkan SwiftShader. All 18 owner/module/window/epoch identity fields remained strict.

reviewed/startup.png, captured at 14:40:00.327Z, shows the red/white Zipper Interactive splash in the original app. Its SHA-256 is 4b2f79af071a4c0298f188b81003d9eedec0cd0c8d0156396f539d2e1f8ef461 (142,332 bytes). The immutable capture and before/after binding are retained. This review did not authorize any dispatched input.

At 14:40:29.444Z the newly written operator command.py supplied 60000 ms to transport.request for a passive menu capture. The retained tools/gltron-logical-fps-local.cjs::commandAck accepts at most 55000 ms and threw bounded ACK required before publishing a numbered command. The wrapper recorded sticky transport-closed.json. This is the exact first blocking control failure, introduced by this worker, not an emulator defect. The subsequent attempted retrieval of nonexistent menu.png also failed and is retained. There was no command replay, closure removal, readiness relaxation, input-handler bypass, guest/config/clock edit or second original attempt. Native events, command queue, input queue and held mouse buttons were empty. No numbered input ACK or independently scheduled UP was needed or claimed.

Because transport was terminal, delegated cleanup used a separate CDP connection only to read passive input state and request Browser.close. Chrome exited normally with code 0 at 14:41:12.092Z. The owning idle driver was then terminated with SIGTERM; both HTTP/debug ports closed, both PIDs were absent, writer lock was absent and its profile was removed. This was external shutdown, not the driver's ordinary quit/cleanup path. No transport gate was bypassed for gameplay.

All 29 unique runtime artifacts (316,424 bytes) were recovered and independently SHA/length verified before normal boat stop. One UUID-keyed POST /sandboxes created bx_ut6m6674; normal stop completed at 14:41:34.262Z, forced:false. Final state is archived, active expiry none (archiveAfter:null); original expiry was 15:39:34.002Z. No old boat was resumed, stopped or deleted. Local one-second disk samples stayed above 3,131,867,136 bytes; allocation, preparation and retrieval guards passed above 2 GB. This is sampled monitoring, not a continuous free-space guarantee.

Validation covered preparation pins, remote pre/post pins, immutable capture hash, evidence recovery and passive empty-input/closed-port cleanup. Source is recorded by exact Git SHA and relevant binary patches; no source/media archive was made. Global evidence checking retains one unrelated Alice containment error. No production code changed, so no build or emulator regression suite was run.

Next separately authorized attempt: use the same published staging recipe with an operator ACK timeout no greater than 55000 ms (the predecessor's tested quit uses 20000), preserving numbered exact ACK and sticky UNKNOWN behavior. This bounded assignment stops at the documented client failure.