Boids and Flip3DTL motion and native close, 18 November 2026

Registered original dx_boids and dx_flip3dtl visibly animate on actual software and WebGL browser executors. All four successful ordinary title closes reach native ExitProcess(0) before host cleanup, then retire guest/render runtimes. Each has separate successful driver, Chrome, Xvfb and openbox actual waits, null signals and both output EOFs.

Evidence: scratch/runs/20261118-dx-boids-flip-current-paired-motion-native-close/.

Original/backend Personally reviewed settled whole-frame motion Completed DRAW_UP / PRESENT Close to native exit
Boids/software Magenta bird moves down/right; cyan terrain bends sweep up/right 9,566→13,802 / 142→205 331 ms
Boids/WebGL Magenta bird turns and moves left/down, blue bird rises/left, red bird appears; terrain changes 6,499→10,251 / 96→152 123 ms
Flip3DTL/software Broad red and blue cube faces become a dominant rotated red face and narrow blue upper face 612→846 / 101→140 76 ms
Flip3DTL/WebGL Red upper and blue lower faces rotate into broad red right and green left faces 1,044→1,422 / 174→237 346 ms

Both selected images per successful scope show settled visible geometry; no startup black-to-scene transition is used as motion. The retained images are entire 1280×1024 private-display XGetImage captures. The browser viewport remains 800×600 without resize events. Both originals naturally select a 640×480 guest display at startup; that backing stays fixed across each pair and before close. The native client is 634×435. No host viewport change, media/configuration substitution or input occurs between the paired observations. Ordinary focus precedes settling and ordinary title close follows review.

Diagnostic comparisons exclude Flip3DTL's changing text: display coordinates x=30..781, y=300..743 contain 2,962 / 4,082 / 60,512 / 60,957 changed scene pixels in table order. These support personal scene review, without a parity or performance interpretation.

The passive observer counts successful returns from the actual render Worker executor. Software reports native software draws and no GL context. WebGL reports webgl1, a live WebGLRenderingContext, 640×480 drawing buffer, attributes and no context loss. Endpoint ID/epoch stay stable and presenter counts advance. All four current scopes use completed PRESENTs; Flip3DTL has zero READBACKs here. Its older readback-only graph is not attributed to this current graph. Renderer/vendor strings are not hardware evidence.

The coherent in-memory graph reuses the tested base Git f6962ba5b1cf2829a66f5cc4d80b293c30719b37, renderer 06f27181f9c977cd1aed99270040efddca00ed75, and owning-Worker unload repair 1ad9afd9201b40a3985421906436bd0023c73ba4. Retained module SHA-256 a6002d5dbcfeb534a0eadd223da616e525aa1542dad2bde381fe54e00ee862f8 is referenced at its existing path without copying or recompiling. Source pins, RPC overlay, bounded native-close transform and observer suffixes reconstruct actual successful HTTP response hashes. Registered inputs and native DLL hashes remain unchanged before/after. Relevant shared WIP is recorded as binary diffs and excluded from served bytes.

Boids handles the ordinary close through native DestroyWindow, PostQuitMessage(0), PeekMessageA returning WM_QUIT, and ExitProcess(0). Flip3DTL traces native DefWindowProcA WM_CLOSE/WM_DESTROY, PostQuitMessage(0), PeekMessageA WM_QUIT and ExitProcess(0). Exit stack arguments are zero before host stop. Retirement and process waits are separate observations. Independent finite supervision finds no leaked tracked children or deadline termination in successful scopes; all tracked runtime PIDs are absent. Forced cleanup is never counted as native close.

Two preparation failures remain separate, permanently closed UNKNOWNs. The first software Boids scope waited on bridge stats.records, which stays zero on the doorbell path although active device records advance. It captures no frames and reaches its finite deadline. A read-only observation before closure proved the counter mismatch; the materially corrected fresh scope uses device records. The first software Flip3DTL scope captures one visible frame, then rejects the current PRESENT route because its assertion expected historical READBACKs. Its retained completed counts prove DRAW_UP=870, PRESENT=144 and READBACK=0. A fresh scope accepts completed PRESENT or READBACK and records the actual route. The still-unlaunched WebGL scope receives these preparation corrections before its first launch. Neither failed endpoint receives requests after closure. The enclosing invoker exits zero even when its supervised driver exits one; actual driver waits are retained and govern these UNKNOWN records. No old UNKNOWN is replayed or retrospectively upgraded.

No product defect was proved and no product code changed, so no new build or broad unchanged qualification was run. The published unload repair and tested capture/close helpers are reused; dd9 dual-executor regression files remain byte-identical. Run-local readiness/route corrections, verification, exact Git/diffs, owned input patches and recovery recipes are retained. Growing transients use owned tmpfs; unique evidence is retrieved and round-trip hashed before removal. Successful runtime /home samples exceed 1,500,000,000 bytes, but a later finalization sample falls to 1,498,841,088 bytes after all originals are closed and retrieved. The task's continuous disk-floor requirement is not met; no further original is launched. Shared capacity recovery is separate from the runtime results.

Foreign WIP and security holds are preserved. No workers, boxes, Wine, hardware/FPS/Windows-parity or deployment claim. Box ID/state/expiry/stop receipt: null / NOT_CREATED / null / null. result.json is written last. Ownership is released and the task stops after publication.