OrganicArt browser adapter recovery, November 7

The ordinary-browser prerequisite is recovered. Current-main OrganicArt reproduces the lower-right white rectangle on both confirmed software and WebGL routes. New native observations corroborate the documented missing logo input; no emulator rendering defect or justified production repair was found. The remaining dependency is authentic Plus! 98 installation metadata/assets matching this package, as described in organic-art-screensavers.md.

Evidence: scratch/runs/20261107-organicart-adapter-recovery/. This task began at 05:04 UTC and finished within its 35-minute bound. No guest configuration, registry override, logo substitute, guest binary, scene asset, crop or hide workaround was introduced. No unknown input was replayed and no held access operation was retried. No agents or workers were launched.

Adapter and route verification

The predecessor required tools/web-input-probe.js as a module, suppressing its guarded main(). The new run-local driver spawns the real CLI script with node --require instrument.cjs tools/web-input-probe.js ..., preserves its terminal status and records browser launch, close, PID and responses. The production CLI API is unchanged.

Prerequisite checks caught and repaired another run-local issue: Puppeteer's module export did not accept direct replacement of launch. The preload now returns a module proxy with the wrapped launch function. A malformed viewport returns nonzero; a deliberately nonexistent app then verifies real CLI main entry, browser launch, meaningful nonzero failure and normal browser close without launching a guest. An instrumented real Chrome smoke page and its capture passed before the original. Earlier failed checks are preserved.

The first original launch, directory software/, actually used WebGL because page initialization overwrote the preload's flag. Its actual GPU counters exposed the mistake; it is excluded from the software results. After that browser closed normally, the confirmed sequence used the supported d3d-renderer=software and d3d-renderer=webgl URL parameters, serially. Both ordinary registered launches used the unchanged /s arguments and exited zero. Their full commands and close receipts are retained.

The first two preliminary layer snapshots also exposed an OffscreenCanvas readout error. The run-local helper was repaired to use convertToBlob() for an existing OffscreenCanvas, with no guest or renderer change. Subsequent snapshots succeeded. Small adapter versions, hashes and patches are retained; these setup defects are not guest failures.

Current original observations

Eight inspected 640×480 scene layers, four per confirmed route, all contain exactly 3,060 white pixels in [550,434,640,468). Software shows animated pink rings, then dense green/red procedural forms. WebGL shows a rotating colored polyhedron, then growing yellow forms against a purple backdrop. Full-page captures also place the rectangle inside the actual scene, separate from the debug toolbar and log. An initial visual impression that software capture 4 lacked the rectangle was wrong; pixel analysis and the full-page capture confirm it remains present. Natural disappearance is supported by the older October 31 evidence, not newly demonstrated here.

Both routes identify scr_oasaver, process 1000, owning window 65537 titled Screen Saver, with the actual displayed 640×480 DirectDraw layer. Surface slots change during scene transitions. Device target slots include software 26 and WebGL 6 then 28; these are not a universal fixed-slot convention. This avoids confusing historical scene slot 6 with blank primary slot 38.

The fresh native observations on both routes record:

The confirmed software route has no GPU object and records 3,687 software immediate draws plus 25,706 execute-triangle trace events by its final snapshot. WebGL's final snapshot records 1,841 draws, 1,451 draw calls, 70,108 triangles, 390 clears, 390 texture clears, four declined clears, zero fallbacks and zero GPU errors. These are diagnostic counters, not FPS measurements.

Focused diagnostics retain full 640×480 device viewport state and clip normalization words, DirectDraw Blt/present/flip boundaries, software vertex coordinates and bounded GPU clear packets. All retained GPU clear rectangles are [0,0,640,480]; no retained clear packet selects the logo rectangle. The rectangle is already in the displayed layer, and software explicitly submits white geometry at those bounds. These observations support the native logo path rather than an invented clear/viewport/presentation repair. Traces retain only the first 24 events per kind and first 48 GPU clear packets; boundary snapshots do not prove completed writes or exclude every unrelated renderer defect. Captures are natural asynchronous observations, not matched frames across routes.

Combined with the earlier native logo-disable/substitute-image experiments, this corroborates the missing authentic input explanation. Those guest-changing experiments were not repeated. The current registered fixture is complete as registered, but it does not supply CALogo.gif or source-verified installation defaults. The intended authentic logo appearance remains unverified.

Provenance and closeout

Source pin: main a03f99823b79dd69ed6f96aa3f106ea96fcbccd8, exact source-file hashes, relevant binary diff and the existing untracked lib/wininit-boot.js patch. Its relevant source hashes equal the predecessor's 59e41298a pin. The locked build passed. All confirmed captures identify actually loaded WASM e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467 (2,182,493 bytes), retained and verified. This differs from the historical October 30 module d83e413f…, whose source-fix reference is de2e338c4 and whose complete compiled-source identity remains unknown. Neither the older 8b227b5c single absent-box capture nor this result establishes a renderer fix.

All 85 registered executable/scene inputs and DLL pins match the predecessor. All 1,161 staged pins were checked before execution and again after, with only explicitly recorded owned adapter revisions excepted. The original executable is 470e28bc…, Microsoft D3DRM is fd80f383…, and D3DXOF is 1f2d0683…; full hashes are in the manifests. The actual Chrome executable is pinned as 22e4d5ef…, version 151.0.7922.108, using cooperative guest execution and ANGLE SwiftShader for WebGL. The installed dependency manifest has 2,306 pins. All 857 captured successful response bodies across the three original launches match staged runtime/fixture hashes, with no unknown successful body or hash mismatch. Some response bodies could not be obtained by CDP, including the software D3DRM request; its successful native load is logged and its served file is pinned unchanged before/after. Do not read the body-observer gaps as verified response hashes. Optional UI/build/API-table endpoints return 404, and initial same-directory D3DXOF probes return 404 before successful loading from the pinned DLL directory. These limitations are preserved in the logs.

All 117 recovered remote artifacts (12,691,269 bytes) passed SHA-256 and size verification before normal shutdown. No source tree/media archive was made. The fresh box bx_c83kuyt5 used a persisted UUID Idempotency-Key, original expiry 05:40:57.897 UTC. Explicitly delegated normal stop was requested once at 05:15:48.518 UTC with forced:false; it completed at 05:15:52.516 UTC. Final state is archived, with no active expiry. The predecessor box was never reused or resumed. All eight recorded browser/controller PIDs, five recorded profiles and owned runtime processes are absent; the four recorded HTTP ports have no listeners. The local messageboard watcher is also closed. Free /home remained above 2 GB at recorded checks.

Only this note is committed; run-local repairs and evidence remain in the run folder. result.json is written last after evidence verification and closeout. Next work requires authentic matching installation metadata/logo input, not a speculative rendering change. No full-scene parity, FPS or historical27 claim.