Diablo II current-main renderer attempt — 2026-11-06

Town movement remains unverified on this version. The one authorized original browser run ended after its first Escape input, with a clean guest ExitProcess(0) and a personally reviewed empty desktop. There was no second original, backend arm, ground click, Den exploration or generated gameplay. No emulator defect or emulator cause was established. This is a harness route failure, not a regression claim about the accepted historical gameplay.

Evidence: scratch/runs/20261106-diablo2-current-backend/. This continues GLD3D-CORPUS-27-20261006, not a new title or a full-27 qualification. Historical town movement, field combat and native checkpoint restoration remain as recorded in the accepted notes; none was reproduced here.

Inputs and transport

The 25 installed files match diablo2-fixtures-20261019.json, totaling 135,378,002 bytes. The unused DiabloIIDemo.exe is absent, so the existing 26-file readiness tool correctly reports one ENOENT. The installed registered diablo2_demo route requires none of those installer bytes. No installation, media recovery download, guest configuration/save/memory edits or source/build changes occurred. The shared browser resolver's existing optional unserved SHELL32 fallback was retained; declared game DLLs were present.

Nine tiny generated transport controls passed. The run-local wrapper reuses tools/gltron-logical-fps-local.cjs's atomic JSON publisher and numbered ACK client. Controls cover temporary partial-file invisibility, rejection of a malformed final file before dispatch, duplicate writer/consumer IDs, ordered ACKs, sequential requests, child-exit cleanup, post-exit rejection and no replay after an unknown ACK. The generated child exited zero and was reaped. These controls prove host transport behavior, not guest input or gameplay. An initial relative-require typo and two preparation assumptions were corrected before any original launch; preparation logs retain the latter failures.

The selected main commit is e2b79dba8fb85cc51cc1593a5fdefff0b68f3688. runtime-pins.json records 589 exact individual runtime inputs; the remote prelaunch hash check and post-run local check both pass. Git plus relevant-source.patch, the untracked lib/wininit-boot.js patch and owned input patches preserve source identity. No source/media archive was created. The existing final production build from the Lords2 worker supplies:

Module SHA-256
standard, actually loaded 6cd858302895de91df5af17e3a3742281e31f369ddb6528d7b96252b3423caac
compat, provisioned only 41b0c704c91ba6fc7e4a6645d9421bf7fd3a23139b0b538b576e1e1b85527dd6

Four live Worker targets, including the owning renderer, expose standard Wasm bytecode matching that hash. The unique module was retrieved. Of 201 actual HTTP responses, all 194 static/bundle successes match pinned bytes, including the normal dev-server index injection and the bundled original renderer DLLs. Four responses are dynamic agent/auth traffic; three are expected 404s for build-info/favicon. Remote media were fully hashed before launch; a second remote post-run media hash was not collected before shutdown. Local media remained unchanged.

Observed original and actual backend

Chrome 151.0.7922.108, V8 15.1.206.16, ran on one fresh separate boat. The page used cross-origin isolation, default guest Workers and the unchanged registered Direct3D selection. Startup shows the authentic Diablo II logo and Blizzard logos in the game canvas. The viewport is 1024×768; startup game buffer is 640×480 with a 684×536 presentation rectangle.

The live render owner is target EC728EA4F0ED24C90BDBEF75A6464CFB, endpoint 1, API legacy, backend webgl. Existing natural-barrier stats advance from 1,721 to 6,932 submitted WebGL draws and fence sequence 1,119 to 2,320. Both completed software counters remain zero, with zero backend errors. The owning Worker's endpoint/control observations agree with the page's endpoint and loaded module. These are actual adapter draw submissions: D3DIMGpu._submit increments draws after t.device.draw, rather than merely reporting a requested renderer or a presentation. They concern startup only. The context reports ANGLE/Mesa Intel UHD Graphics 620; physical GPU completion, GPU timing and town consumption remain unverified.

Startup pixels were captured at approximately 22:48:22 UTC. The first ordinary Escape was submitted at approximately 22:48:47 UTC, after a state/owner read. That preinput read has no pixels; the reviewed startup state is 24.945 seconds older. The transport guard checked the capture hash and review reference but did not establish that the animated guest remained on the same screen. At 22:48:47.860 UTC the log records ExitProcess code: 0; the 22:48:52 capture shows the empty desktop and no running app. Escape may have reached a later menu, but the exact screen at input is unknown. No crash, trap or missing API was observed. A stale scene review is the demonstrated harness limitation; assigning an emulator cause would exceed the evidence.

Cleanup and recovery

Original start: 22:48:04.164 UTC. Chrome exited zero without a signal at 22:49:08.483 UTC. Unique logs/captures/module were retrieved by 22:49:33; the recorded controller/Chrome PIDs were absent, both private ports refused connections, and the browser profile and writer lock were gone. No shared local runtime locks, ports, browser or source/build files were owned.

Fresh boat bx_8vanavxu was created with a persisted UUID Idempotency-Key, initial expiry 23:45:19.723 UTC. Delegated normal stop stop_c183402b1d198c0e completed at 22:49:50.572 UTC, forced=false, error=null; API state is archived, stopped at 22:49:50.552 UTC, active expiry null. Total execution, review, retrieval and stop took under two minutes, within the 35-minute limit. No access403, automated flag or lifecycle retry occurred. Local free disk remained above 2 GB.

Recover from the recorded Git commit plus exact patches and the existing path/hash media references in runtime-pins.json; retain the recovered unique module and both reviewed captures. recovery.txt records the detailed recipe. No production fix or build was warranted, and no second original was attempted. Next work requires a separately authorized original with state-driven menu admission and fresh scene review before input, then the requested reviewed before/idle/ground-click/settled town captures. Audio, FPS, native Windows, software-backend gameplay and full-corpus completion are not qualified here.