OrganicArt natural motion: WebGL pair reviewed; SOFTWARE startup closes

Run scratch/runs/20261118-oasaver-paired-natural-motion/ consumes exactly the two newly authorized local originals, one SOFTWARE and one WebGL. One meaningful WebGL motion pair is personally reviewed. The SOFTWARE original closes at its unchanged 15-second binding guard before a frame is admitted; the requested two-backend result remains blocked. No replay, guard extension, behavior repair or additional original is performed.

Reviewed natural WebGL motion

The registered scr_oasaver, original OASAVER.SCR /s, reaches binding in 1,332 ms. webgl/natural-scene-before.png and webgl/natural-scene-after.png are full lossless 1280×960 page frames. Both already show the same cloud-backed scene with silver curved and star-like objects. Orbiting objects move and rotate; the central cluster changes orientation. This is scene motion, not black-to-first-scene, a cursor or UI change. The known lower-right white rectangle remains in both images; its authentic-logo correctness remains unresolved.

The actual D3D render Worker executes GpuStreamWebGL: completed observer PRESENT counts advance 43 → 101, DRAW_UP 1,117 → 2,636. The software executor remains unused. Its actual context reports WebGL 1.0 and ANGLE (Intel, Mesa Intel(R) UHD Graphics 620 (KBL GT2), OpenGL 4.6 (Core Profile) Mesa 23.2.1); launch flags alone are not used to identify the renderer. Deferred module inventory after the complete motion interval finds the actual page, guest Worker and render Worker modules all equal to the pinned module hash.

Natural requested waits are 1,000 ms before the first selected frame and 2,000 ms before the second. Post-capture markers are 2,325 ms apart; this includes capture/observer work and is not an exact exposure interval or FPS measurement. Individual sleep durations were not separately instrumented. Offline comparison finds 17,043 changed pixels bounded by x333–907, y115–841, entirely below the unchanged top UI and away from the white box. Pixels supplement the personal whole-image review rather than define it.

Both captures preserve the bound native identity: app process 1000, Screen Saver HWND 65537/generation 1, guest endpoint 1, render endpoint 1 and worker epoch 1, with unchanged 640×480 guest geometry and 1280×960 viewport. The original 300-ms stability, no-resize inspection and postbind guards pass. No input is recorded. Eleven passive startup samples separately show initial PRESENT 0 with composition, then positive PRESENT; native execution advances. Focus comes from ordinary slice replies with unknown age; modal HWND, API and loading-completion flag remain unknown. No reservation guard is waived.

SOFTWARE blocker and limits

The first authorized original uses the ordinary helper readiness conjunction, without the opt-in startup sampler. At bind +15,015 ms it returns pre-binding stabilization timeout, state UNKNOWN_CLOSED. The retained observation has one visible Screen Saver owned by process 1000, with native guest/render identities, and ready:false. Independent PRESENT/composition values, actual executor progress and instantiated module bytes are unavailable for this closed run. HTTP receipts establish the served module and graph, not actual module-instance inventory. No screenshot, motion pair or zero-PRESENT claim is made. No page exception was recorded, which does not identify the cause. No endpoint is queried after closure.

The second authorized original enables the existing published passive startup observer to retain separate predicates. This option difference is explicit in webgl-preparation.json and prevents any observer-equivalence inference from these two originals. The prior published native equivalence proof is reused; no new behavior-equivalence claim is needed. The historical 15-second failure still has unknown cause. This new SOFTWARE timeout supplies no causal repair or retrospective diagnosis. All originals are consumed and closed; no held investigation is left running.

Exact graph, proof reuse and cleanup

Both originals serve the accepted coherent published Git graph 47095dcb6c84178f63dde8bb61e9fd142aa1f4b9, explicitly an older pin and not current main. The existing module is eb2395e1c115980dd78dbb646a26f755a4119a2a1e086942e715240e07c34bc6 (2,408,519 bytes), retained once in the preceding bounded-startup run and referenced by recovery recipe. All 88 original registered media/DLL files match before/after. Exact Git-served source hashes, the small additive public executor-observer suffixes and every HTTP response hash are retained. Shared working-tree differences are recorded as binary patches and never applied.

The published 47095dcb6 observer's separate-predicate native controls, ordinary/observed equivalence, focused tests and locked build from 20261118-oasaver-bounded-startup-progress-cause are reused. Production tools and tests are read-only; there is no runtime behavior change and therefore no new behavior old-fail/new-pass or build claim. Run-local JavaScript prepares, supervises and validates evidence. Immutable graph/media/module checks and a native PNG construction roundtrip precede runtime clocks. An initial offline preparation assertion compared compiler-normalized WAT hashes with raw served Git blobs; it was corrected to compare the successful original's actual pins-before.json graph before either runtime started. No original was spent on that failed preparation.

Each runtime is owned through tools/native-supervise.cjs, with a separate finite watchdog process independently sampling free home disk. SOFTWARE native and command waits are exit 1; WebGL waits are exit 0. Both watchdogs are canceled and actually waited at exit 0; all four command log streams reach actual EOF. No native process leaks. Offline cleanup verifies all 55 recorded process IDs nonlive and both known HTTP listeners absent. Watchdog minimum free disk is 3,109,085,184 bytes for SOFTWARE and 2,775,576,576 for WebGL, above 1.5 GB; no RAM gate is used. The first board watcher has actual tool wait 143 with unknown signal sender, explicitly retained in the final evidence. No box is created: ID null, state NOT_CREATED, expiry null, stop receipt null.

No source snapshot, media copy, frame series, capture bundle, Wine, Python, worker, deployment or foreign-owner edit is introduced. The two selected whole PNGs are the only retained image frames. Full27, backend scene parity, FPS, audio, authentic-logo correctness and ordinary native Quit remain open. Artifacts, readback validation and cleanup precede result.json, written last. The report is committed and immediately strict pulled/pushed, then ownership is released and work stops.