OrganicArt SOFTWARE presentation path: motion observed; old cause unresolved

Run scratch/runs/20261118-oasaver-software-presentation-path-cause/ publishes bounded observer 054baf9e5 and consumes one authorized causal SOFTWARE original. It binds in 4,110 ms and yields a personally reviewed natural-motion pair. The preceding 15-second colored-image/zero-stream-PRESENT failure does not recur. Its exact pixel provenance and cause remain UNKNOWN; no behavior remedy is proved and no conditional second original is run. The new pair is an observed success without a repair, not an after-remedy causal result.

Actual presentation, rather than backend intent

The fresh registered scr_oasaver, original OASAVER.SCR /s, has two distinct presentation stages:

Stage Actual evidence
Before stream draws Native dx_trace kind 5, surface slot 5, 16 bpp; canonical surface 2097157, 640×480 RGB565, successfully attached/uploaded to HWND 65537. The page cache records directDraw:true, version 2, flush count 1, no pending guest upload, active back canvas 2097157, compositor publication 9. Actual render executor contains only two LAYOUT records.
Scene rendering Actual render Worker GpuStreamSoftware executes DRAW_UP/PRESENT. Window layer becomes d3dim-gps, kind gpu, with advancing presentation sequence; compositor publications advance. The WebGL executor remains unused.
Selected natural pair Software completed PRESENT 21 → 33, DRAW_UP 124 → 196; same process 1000, HWND 65537/generation 1, guest endpoint 1, render endpoint 1 and epoch 1. Three actual module-byte inventories all match eb239….

Thus bridge.presenter.presents counts the stream presenter; zero is not a count of all DirectDraw/GDI presentation. This execution directly establishes a canonical DirectDraw presentation before stream rendering. It does not prove that its initial upload contains rendered geometry, or retrospectively identify the source of the old colored fan-shaped scene. That closed predecessor lacks actual executor inventory, active canonical target and publication provenance. Its screenshot is neither reclassified as a D3DIM frame nor dismissed as blank, static UI, a bitmap, or the wrong owner. Those alternatives remain unseparated for that historical image. The new native static-GDI control demonstrates why an upload and composition alone also cannot admit a completed D3DIM scene.

The first sampled presenter-positive value is 2 at request-start bind +2,025 ms. Fourteen startup observations retain native EIP/yield progress, file-import returns, per-realm executor counts, canonical targets and publications. Initial yield 5 becomes 0; the initial EIP is 0x747f4a96. File-read-result returns advance, but no generic loading-completion flag or current native API is invented. Host-import returns are not guest-API completion receipts. Page and guest observations include forwarded calls and must not be added together. Getter, page-cache and Worker reads are sequential, not atomic snapshots or an exact first-frame clock.

Reviewed motion and exact inputs

original/natural-scene-before.png and natural-scene-after.png are complete lossless 1280×960 captures. Both show the same cloud-backed scene. The left silver/blue and right dark articulated objects rotate/change shape. The top UI, background and known lower-right white rectangle remain unchanged. Offline comparison finds 14,703 changed pixels inside x173–1107/y289–706, supporting the whole-image review. Requested ordinary sleeps are 1,000 and 2,000 ms; post-capture markers are 2,676 ms apart, including capture/observation work. Individual sleep durations and exact exposure times were not measured. This is not an FPS, parity or authentic-logo result.

All 1,386 source pins and 88 registered media/DLL pins match the successful 47095 startup and the observed timeout. Their launchers use fresh Chromium contexts without carried browser profiles or registry overrides. Equal static inputs do not establish identical native random/clock/scene state. Media requests are retained separately from actual native file activity. An offline GIF-content audit records only in-memory thumbnail statistics, not an image identity or pixel-origin proof; no media copy is retained.

The coherent base graph is explicitly older published Git 47095dcb6c84178f63dde8bb61e9fd142aa1f4b9, with additive, recorded observer prefixes/suffixes from published 054baf9e5. Actual page, guest and render module inventories are all eb2395e1c115980dd78dbb646a26f755a4119a2a1e086942e715240e07c34bc6 (2,408,519 bytes), retained once in the preceding startup run. Debugger runs only after the complete motion interval. Before then, actual instantiation metadata explicitly leaves compiled-module byte identity null; import/export counts are not hashes. Relevant current working-tree differences are recorded as unapplied binary patches. Git remains the source archive.

The accepted 37118 WebGL pair and its hashes are preserved without rerunning it. The newly witnessed SOFTWARE pair does not explain why the earlier SOFTWARE original stayed at two device records for 15 seconds.

Observer proof and cleanup

The opt-in observer wraps existing imports and instantiations, adding no guest calls, API tracing flags, input, repaint, viewport changes or pacing. Native ordinary/observed controls produce identical whole-image pixels, no input and no page errors. Final native negatives retain static GDI UI with zero stream PRESENT (15,004-ms rejection), completed PRESENT with a hidden window, and an actual second HWND after binding (rejected before capture). Offline mutations of the retained native baseline reject wrong process, endpoint, epoch, HWND and generation before capture; these are explicitly offline tests, not native process/epoch replacements. The default-off zero control lacks path evidence; the opt-in equivalent supplies it while still rejecting readiness. This is an observation-capability improvement, not an old-fail/new-pass behavior repair.

Early authored controls caught a missing page-installer argument, corrected before any original. Strict-mode preservation and bounded metadata were also finalized before publication. Superseded attempts, errors and reductions are recorded. Focused tests and the locked full build pass. A shared no-Python gate failure was repaired by its existing coordinator owner; this worker did not edit that owner's files or the allowlist. The observer is committed and strict pulled/pushed before original preparation and launch.

Every native run uses tools/native-supervise.cjs, an independent finite watchdog and actual free-home sampling, with worker-owned cleanup and no RAM gate. The original native/command/browser/watchdog actual waits are exit 0, both command streams reach EOF and no native leak is reported. Original minimum free home is 2,472,087,552 bytes. Final receipts verify owned processes nonlive and known listeners absent. No endpoint is revisited after a closed UNKNOWN; the successful original is also fully closed. Artifacts, exact Git, relevant and owned patches, hash readbacks and cleanup precede result.json, written last. Box ID null, state NOT_CREATED, expiry null, stop receipt null.

The concrete limit is the unreproduced historical cause and absent causal remedy, not a newly asserted rendering failure. No second original, replay, WebGL rerun, guard relaxation, worker, box, Wine, Python, deployment, source archive, capture bundle or media copy is used. Full27, parity, FPS, audio, authentic-logo correctness and ordinary native Quit remain open. Ownership is released and work stops.