GTA2 original direct launch, software walking and driving — 2026-11-07

Settled walking, native Enter vehicle entry and separate settled driving were personally reviewed on the original GTA2 Wild Demo using the production SOFTWARE renderer. This establishes the narrow software gameplay cell for GLD3D-CORPUS-27-20261006. WebGL evidence 27fdd86e9 remains separate. No emulator defect was demonstrated and no production code changed. There is no combat, audio, FPS, parity or full27 claim.

Evidence: scratch/runs/20261107-gta2-direct-software/.

Before allocating, seven controls ran the actual production direct initializer extracted from index.html. They cover the filtered-out original and dx_boids, delayed registry restoration with DOMContentLoaded, wrong ID, cancellation, a present option, sign-in resume, and explicit shell.launchApp(wanted, {directIntent, startedAt:0}) arguments. The MCM direct-route harness was reused with its production behavior preserved. The browser used ?debug&threads&d3d-renderer=software&app=ID; the dropdown remained Notepad. No catalog option or fake launch was introduced.

The literal action controls execute the actual browser harness executor with mock transport. The intended arrows, Enter, Escape and 50/100/2000 ms holds, balanced clicks/double-clicks, capture, quit and invalid key/wait checks passed. No text entry was intended. Initial preparatory assertions incorrectly counted capture as a dispatch and allowed an earlier mock Alt tap to leave menu mode active for Escape; both control expectations were corrected before allocation. The production renderer/DirectInput key state checks pass independently. Durations 0, 49, 2001, 3500, -1, NaN and 100.5 reject before transport and input; accepted bounds are 50–2000 ms. Each Chrome session also recorded five balanced trusted arrow/Enter pairs on about:blank before selecting an app, with zero events from invalid-duration requests.

The dependency inventory includes tools/protected-writer-entries.js, every render-artifact.js identity input and the static relative require closure. Miles resolves at its actual app-local URL: /binaries/candidates/gta2-demo/installed/Program_Executable_Files/mss32.dll. Both Miles and COMCTL32 returned HTTP 200 with pinned bytes before each browser launch. No /null request occurred. All 694 staged pins match before and after. The retained supplied media comprises 69 registered paths plus the installed EXE and Miles DLL: 71 files, 26,207,806 bytes, unchanged before/after. The older 73-payload distribution and upstream provenance are not reconstructed.

Installed file Bytes SHA-256
gta2.exe 2,893,824 97ad743b6ec9ea1be95282053c3127084acd764a7f37a3e57e2268af52d02ae3
mss32.dll 331,776 0974b244354a5d13e0711db15430c05f7949dc279b63897146f304d1401153fc

Before the original, the generated x86 public COM/vtable Surface4/7 regression passed remotely through its prepared and consumed com-preflight.wasm and identity manifest. The real dx_boids guest then launched in a Worker through the production direct URL. Its owning renderer Worker target 4D59EDE7FFECCA3CEA2B965BA90A9FA5 had endpoint 1, legacy/software, no adapter (therefore no GPU adapter), no endpoint failure and control error 0. Direct software draw completions advanced 195,506 → 222,507 and completed sequence 52,208 → 59,108; guest fallback count was 0. The software path deliberately has no adapter and counts draws in endpoint evidence; an inherited WebGL adapter assertion was corrected by reading this production contract before allocation. The sample's colored boids and boundaries were personally viewed.

A separate generated guest called real COMCTL32 PropertySheetA. Its ordinary property sheet and label were personally viewed; a balanced native click on OK at page coordinates 173,418 dismissed it, also personally viewed. This used a scratch generated fixture entry and explicit shell launch without changing the dropdown or original registry entry. Both generated browsers closed normally, and their processes, ports, profiles and writer locks were absent before the original. The earlier software attempt never launched a guest and remains a launch-readiness result, not evidence of a renderer defect.

Exactly one original started at 02:29:26.512 UTC. Each input used reviewed geometry and the existing freshness admission protocol. The capture sequence establishes:

All 20 original commands have complete ACKs, including quit. Three requests (play, walk5, entry1) were refused stale before dispatch with explicit effects 0, then refreshed under new IDs. No UNKNOWN/partial input or replay occurred. A mistaken read-only fetch requested the nonexistent play capture after the first refusal; it failed without guest effect and is retained as first-error.json. Subsequent request handling checked the ACK before fetching. The cleanup template's literal UNKNOWN; never replay is a generic policy label, not an input outcome.

The actual original renderer Worker target AEA95198DB365AE713397F098603E706, session 935537845319E649A016CEB18F2597BB, owns endpoint 1, legacy/software. At the final settled capture it records 345,907 direct software draw completions, completed sequence 612,206, batches 333,054, commands 345,907, ready 1, error 0, failure null and adapter absent. The guest reports fallback count 0 and null GL renderer/version. These identify software ownership, not FPS. The browser presented screenshots supply visual evidence.

All three executing original Workers (main guest, guest thread and renderer) yielded actual Wasm bytecode with SHA-256 f18229189ae8e50fe938baaa6f2fe2821ee60c297acf7116f5cfeb8529915eaa (2,181,162 bytes), matching the staged standard module. All 29 executing JavaScript source records match their exact staged pins. Chrome was 151.0.7922.108, V8 15.1.206.16. Source staging HEAD was cb5f15aa7ae1182e68e38efc46c5923140d52e2c; relevant dirty and new-file patches are retained. Existing modules were reused; fresh build provenance is not inferred from that source HEAD. No local guest runtime or build was run.

All 250 remote artifacts were retrieved or verified against identical local bytes, with repeated modules deduplicated and mapped. All 601 complete HTTP 200 responses match exact files; no successful response is unmapped. The audit separately retains non-200 build-info, auth, favicon, sample-icon and optional font-bundle responses; individual font loads succeeded. URL decoding was needed to map the successful gta2%20manager.exe media response. App-local SHELL32/OLE32 used the recorded built-in stub fallback; no replacement DLL was introduced. Guest media, config and save inputs were not edited.

All three Chrome exits were 0. Processes, six ports, profiles and writer locks are absent. The immediate terminal inventory caught the original controller finishing exit; final-original-process-check-remote-result.json confirms it absent afterward. Delegated normal stop stop_0c1d0d440052ef3a completed at 02:34:14.783 UTC, forced false, error null. Boat bx_tbyx72n9 is archived, active expiry null. Allocation through retrieval and completed stop took about 6 minutes 51 seconds, below 35 minutes. Local disk stayed above 2 GB. No automated risk flag, old403 resume, held-risk retry, second original, nested worker, source/media archive or deployment occurred.

Findings only are published; no production repair or full build is claimed. The requested software walking, native entry and settled driving evidence is complete. Other gameplay dimensions require separate evidence.