Viewer supplied-model motion and ordinary native close, 2026-11-18

Registered dx_viewer visibly rotates its original supplied Microsoft model on both actual software and WebGL browser executors. Ordinary title close reaches native PostQuitMessage(0), PeekMessageA returning WM_QUIT, and ExitProcess(0), followed by guest/render retirement and individual successful driver, Chrome, Xvfb and openbox waits. Evidence: scratch/runs/20261118-dx-viewer-current-browser-paired-motion-close/.

Scope Personally inspected whole-frame motion Actual completed DRAW_UP / PRESENT Native exit after close
software Readable broad face to narrow reversed side; 18,240 changed scene pixels 207→431 / 103→214 99 ms, code 0
webgl-active-observer Mirrored rear word to readable front/oblique word; 9,758 changed scene pixels 495→1,094 / 246→545 220 ms, code 0

The scene pixel counts use the same 376×354 client area in retained whole 1280×1024 display images; no cropped evidence is substituted. The page and native canvas remain 800×600 from startup through both observations, with zero observed resize events. The initial ordinary focus click precedes both images; input records are identical across each pair. No model, media, executable or app config was edited. Ordinary startup loads the model; no File/Open, injected mesh, private guest call, clock change or forced pump was needed.

The actual software executor reports completed native software draws and presents with no WebGL context. The actual WebGL executor reports completed draws/presents, backend webgl1, a live WebGLRenderingContext instance, 376×354 drawing buffer, context attributes and no context loss. Bridge endpoint ID/epoch remain stable and presenter counts increase. Executor observations wrap the pinned production execute methods and count only successful returns. The renderer vendor string is retained as untrusted metadata; it establishes no hardware claim. There is no performance, hardware or Windows-parity claim.

Prepared graph and reused repair

Both successful originals use Git runtime graph f6962ba5b1cf2829a66f5cc4d80b293c30719b37, renderer 06f27181f9c977cd1aed99270040efddca00ed75, and published owning-Worker unload repair 1ad9afd9201b40a3985421906436bd0023c73ba4. The existing module is SHA-256 a6002d5dbcfeb534a0eadd223da616e525aa1542dad2bde381fe54e00ee862f8, referenced at its retained path without copying or recompiling it. Preparation constructs the source graph in memory before launching each finite scope. Source pins, explicit RPC overlay, bounded native API diagnostic and passive observer suffix hashes reproduce the actual HTTP bytes. Shared foreign WIP is recorded as binary diffs and is not served.

All registered Viewer files and native d3drm.dll, d3dxof.dll and oleaut32.dll are referenced from existing paths with before/after hashes. The loader's app-local d3dxof.dll probe returns 404 before resolving the served common DLL; the complete HTTP list also retains absent optional shell resources. These do not prevent the observed model or clean exit.

The tested Globe/Tunnel close route and the published dd9 dual-executor regression are reused. Both dd9 regression files remain byte-identical; no broad unchanged qualification was run. No product defect was proved or product code changed, so no new build was needed. The old retained-rendering note's Wine citation is not used as ground truth.

Closed observer failure and corrected scope

The first new WebGL scope (webgl/) captured one whole image, then failed the executor assertion: installing the unused software observer overwrote the global record with an empty software record, although WebGL was executing. It remains permanently UNKNOWN/closed, with driver exit 1, explicit failed-scope Chrome cleanup and no leaked supervised children. It is not native-close evidence. No later endpoint request was made against it.

The materially changed webgl-active-observer/ publishes the observation from the actual successful execute call. Its paired images and context/executor records pass, then ordinary close completes. CDP inspection creates/detaches Worker sessions only; no second Puppeteer browser attachment or viewport setter is used. All earlier UNKNOWNs, including the old software Viewer scope affected by implicit inspection resizing, stay closed.

Native completion, waits and retention

Both successful originals record the native API entry for ExitProcess with stack argument 0 before browser cleanup. Native quit-message observations and guest/render retirement are verified separately from host process bookkeeping. Each driver/supervisor, Chrome, Xvfb and openbox has its own actual wait with code 0, null signal and both output EOFs. Independent finite supervision reports no deadline termination or leaked children; all tracked runtime PIDs are absent. The original bound is 150 seconds, close bound 20 seconds, and independent supervisor bound 165 seconds plus 15 seconds for cleanup. No forced cleanup is counted as successful native close.

Unique evidence is retained and verified before owned tmpfs removal. The private display socket, lock and authentication file are gone. Minimum sampled /home free bytes are 1,549,287,424 (software) and 1,530,384,384 (successful WebGL); these 100 ms samples are not a continuous-space claim. Only owned closed metadata was losslessly compressed, with round-trip and SHA checks recorded before unlink.

verification.json checks source/media/HTTP/module hashes, actual executor and presenter progress, stable viewport/input, scene changes, native quit/exit, individual waits and absent PIDs. Recovery uses Git plus relevant binary diffs, owned run-input patches and the existing unique retained module. No source-tree snapshot, frame series, media copy, Wine reference, public deployment, worker or box was created. Box ID/state/expiry/stop receipt: null / NOT_CREATED / null / null. Foreign WIP and security holds are preserved. result.json is written last.