Original Deus Ex owning inspection and natural surface evidence

The repaired owning-instance inspector reached native readiness and the original natural Draw/Unlock/return pair. The fresh host PNG was personally reviewed: the game area is black. The first runtime failure was afterward, while exporting the surface PNG: TypeError: Illegal invocation at 09:19:40.042Z. The one browser closed normally at 09:19:40.090Z. No second browser or reload was attempted. Surface PNG review remains unavailable; this assignment ends with the exact first failure and its tested offline helper repair.

I read the published cb0c5fa1f, cbcfd14c1 and c4ea72e5b notes before using the predecessor's owner-inspect.cjs and next-driver-fixed.cjs. Evidence and runnable small helpers are retained in scratch/runs/20261104-deusex-owner-filter-natural/.

Natural execution and raw identities

Chrome/151.0.7922.108, PID 77813, ran 09:19:13.384Z–09:19:40.090Z. The actual original EXE was measured. Slot 0 / thread 1 was checked before Engine module scanning; nonowners remain ineligible and owning mapping errors remain fatal. At 09:19:20.899Z, native GIsRunning=1 and the visible, source-bearing 640×480 HWND 0x10003 passed the unchanged readiness predicate. The collector started at 09:19:21.017Z, owning endpoint identity 1.

The single authorized extended observation requested exactly 10,000,000 blocks / 120,000 ms and completed naturally after 1,121,609 blocks / 18,419.070000000298 ms. Its frozen terminal is render-route-return-candidate-unverified-lifetime, at saved caller 0x1134426e, ESP 0x179ff88c, EBP 0x179ff8b0, thread 1, yield 0, halt 1. It records 393 events: all nine original checkpoints and 192 matching Render ESP save/restore pairs. Original signatures, module generations, ESP ownership, terminal limits and frozen stop replies are verified by analyze.cjs.

All nine raw identityObservations survive in collector status, assessment, collector stop and final stop. Every identity observation reports no diagnostic errors and stable modules. Raw fields include Engine 0x103ef0a0, viewport 0x19309380, Engine.Render 0x1046ee00 and player 0x10384600. Reciprocal UObject membership and unchanged class/slot bytes are boundary observations, not allocation lifetime evidence. The legacy top-level null identity fields were not substituted for or used to discard the raw records. Lifetime and full viewport-to-host binding qualification remain unknown.

Observed surface and actual page imports

The natural route is DirectDraw (viewport flags 9), not the DIB branch. At natural Lock return and Unlock call the viewport pixel storage is WASM 475574272, 640×480, 16 bpp, stride 1280. Unlock ECX equals the viewport; its original presentation argument is 1, HWND/present HWND 0x10003. The locked back-surface candidate is DX slot 2 and its storage matches there. After Unlock, slot 2 points to a different back buffer; that difference is reported rather than called a renderer defect.

The page observer retained 13 actual import observations, with no observer errors or event-limit exhaustion. Its final sequence records:

Serial Actual import Evidence
9 dx_trace(6,1,2,475574272,474935296) Native route's observed storage rotation
10 dx_trace(5,1,16,475574272,0) Observed presentation trace
11 gdi_surface_create(2097153,640,480,16,475574272,1280,1,0,0,63488,2016,31) Returns 1; actual canonical presentation storage updated
12 gdi_surface_attach(2097153,65539) Returns 1; target HWND 0x10003
13 gdi_surface_upload(2097153,0,0,640,480) Returns 1; write sequence 36, version 6

After serial 13, presentation ID 0x200001 has bitsWa=storageOffset=475574272 in the actual shared guest memory, matching the natural Unlock source. Its canvas identity 12 equals both the target window's attached back canvas and DirectDraw owner canvas. The source canonical 614,400 bytes have zero nonzero bytes. The derived RGBA cache has only 307,200 nonzero bytes (opaque black alpha), consistent with the personally reviewed black host image. The ordinary host flush count advances from 2 at upload to 3 at the read afterward; the observer invokes no flush. The separate 648×488 GDI window surface is recorded and is not the attached DirectDraw canvas.

These are actual storage/upload/target/canvas observations during the natural interval. They do not supply a native publication EIP or an allocation serial, and temporal association alone does not prove object lifetime or fully qualify binding. No production publication mismatch is demonstrated. No speculative renderer change was made.

Exact failure and tested helper repair

The surface PNG helper borrowed HTMLCanvasElement.prototype.toDataURL.call(p.canvas, 'image/png'). The host surface factory in lib/host-imports.js preferentially creates an OffscreenCanvas. The HTML encoder rejects that canvas brand, explaining the recorded illegal invocation. This is a diagnostic helper defect, not proof that the renderer failed to display nonblack pixels.

The repaired run-owned page-observation.cjs makes png asynchronous. It uses OffscreenCanvas.prototype.convertToBlob.call(canvas, {type:'image/png'}) plus FileReader for an OffscreenCanvas, and the native HTML prototype encoder only for an HTMLCanvasElement. Missing/mismatched HWND and stale attached canvas return null; unsupported canvas brands and encoder failures remain errors. Neither path invokes the instance's lazy encoder or a guest export, flush, repaint, upload, input or memory write.

next-driver-fixed.cjs now writes the complete raw result before attempting PNG export. The browser-tested namespace stays immutable. Only separate owner-filter-natural-offline-fix inputs contain the repair. test-png-brand.cjs reproduces the predecessor failure with strict modeled brand checks and passes both corrected native-prototype routes, wrong HWND, stale canvas, unsupported brand, native encoder rejection, and zero lazy flush or guest writes. These are modeled encoders, not a second browser validation. The full actual WASM/owning Worker/instruction/next-driver suite also passes with the repair. A further full actual Worker driver regression injects an encoder exception and checks that the raw result, final stop and browser cleanup survive. Small helper patches and tests are retained in evidence.

Before Chrome, all 511 source/helper/module/media pins matched. Actual WASM/original EXE+Engine/Core/WinDrv/Render fixtures, owning Draw/ESP/native Worker/full next driver, owner-filter positive/negative tests, raw correlation and wrong storage/HWND/stale canvas controls passed. Actual Puppeteer 25.7.0 CJS/ESM imports, driver closure, generated region map, cache graph (98 page / 13 Worker scripts), all 144 served resource SHA/isolation/no-store checks, selector negatives and unchanged readiness fixtures passed before launch. The qualified manifest SHA is recorded in prebrowser-pass.json.

The locked full build passed after closure. This was an offline shared-main build including another worker's D3D changes; its newly generated module was never transferred to or used by this browser. No production or repository test source was edited by this task. Only this note is committed; diagnostic helper fixes/tests remain run-specific, with exact owned new-file patches.

Capture, recovery and custody

I personally inspected recovered/fresh-host.png, 93,534 bytes, SHA-256 f12738a54f7f61a8784dade915f2e1a14457d0437b8427c0a3d1ca91859601cd. It shows Deus Ex Demo selected, original launch logs, and a black game area; there is no visible gameplay claim. No fresh surface PNG exists because of the first failure. No replacement surface image was reconstructed or mislabeled.

The closed browser namespace contains 82 artifacts / 5,340,377 bytes, recovered or matched to identical small local inputs with full SHA/size verification, snapshot 09:19:53.476251Z. Raw terminal recovery is exact from final-stop.json, not a later native-PC substitute. The separate offline repair has its own closed SHA manifest and checks. All original 511 pins remain unchanged, Chrome/driver are absent, and HTTP 8123 is closed.

Git source provenance is 0b7318794f8dbf8dc5fdd552d30ab2e97b76480d, relevant binary diff, owned new helper/note patches and the final local artifact seal. The actual retained WASM SHA-256 is 129a71a01a5fbc8b07d6fa242a5a3ebc6a29f5aed9fb4b6a47d9318a764c1f52. Media and both retained module paths/full SHA are in retained-input-path-sha.json; none were copied. No source archive was made.

Handoff: bx_3k2ukgfj is left RUNNING (API idle, healthy/hydrated), expiry 2026-11-04T09:55:33.875Z. Coordinator retains shutdown ownership and receives custody. No lifecycle, allocation, source-box access, local heavy browser, subworker, successor, Heroes II or deployment action occurred. The extended profile has now been consumed once. Any further runtime validation of the encoder repair requires a separately authorized fresh observation. This single task stops after publication.