Boids visible-flock recovery, 7 November 2026

The interrupted preparation was recovered without duplicating an allocation or an original launch. The fresh Worker observation shows two moving colored birds and cyan terrain at 30, 60 and 120 seconds. The serial cooperative run shows moving terrain without birds at those checkpoints. All retained complete geometry frames submit all 13 birds: the Worker has 11 wholly outside the homogeneous frustum and two intersecting it; cooperative has all 13 outside. Every submitted bird position, depth and reciprocal W matches the raw renderer input exactly. A clustered visible flock and correctness of the upstream model transformation remain unqualified. No emulator defect or fix is demonstrated.

Evidence: scratch/runs/20261107-boids-visible-flock/.

Interruption and preflight recovery

The predecessor launcher 1940025 and child exited 143 at 10:43:47 UTC, according to the recovery assignment. It produced neither a final response nor result. Its local fixture had already captured scene15/30/45 and failed its nonempty geometry assertion at 10:43:42.395. This failure preceded the launcher interruption; it was not caused by the interruption. allocate.py and launch.py were prepared but never executed. Event inspection, absent allocation/lifecycle/provision files and the live box inventory agree. The old board watcher, fixture parent and Chrome PIDs were absent; both ports were closed and the private profile removed. BG2 and Heretic were untouched.

The observer had attached to the render endpoint's D3DIMRenderWorker wrapper, which has no _draw, _clear or stats. Its natural GPU hooks belong to the underlying D3DIMGpu. The run-local repair uses the existing diagnostic factory attachment to reach that exact GPU owner, checks the methods, implements the required receipt method, and retains a bounded rolling 512-draw history. It adds no draw, flush, readback, fence, guest scene write or GL retirement work.

The exact Worker driver then passed generated public-COM production GuestWorker/render-Worker execution, scalar readiness controls and final full-compositor captures at 15/30/45 seconds. It observed 25 raw draws and 25 completed submissions, with translated production image metadata. The initial recovery attempt also exposed the wrapper's required receipt method; that failure is preserved. The focused draw-trace test passed. All 315 relevant source pins reconstruct exactly from the recorded Git SHA plus patches.

Worker geometry and pixels

One UUID-idempotent box, bx_z65jbe9z, was created at 10:48:27.947 UTC. Its immutable deadline is 11:23:27.605 UTC and initial service expiry 11:23:27.788 UTC. The original Worker launch ran once, with native 640×480 render target and the unchanged full 1024×712 compositor capture. No gameplay input, camera, clock, original configuration or viewport override was used. The driver observes natural wall-clock checkpoints, without pinning guest time.

recovered/case-worker/scene-{30,60,120}.png were personally inspected. Two separated purple/red triangular birds move over cyan terrain boundaries; no clustered flock is established. The owning backend reports WebGL 2 on ANGLE Vulkan SwiftShader (Subzero), forced by the provider Chrome wrapper; /dev/dri is absent. This is not physical-GPU evidence.

Each geometry checkpoint retains seven complete 67-draw frames: 50 terrain strips, 13 birds of ten triangles each, and four obstacle draws. Each bird batch's 13 source spans and 130 completed triangles preserve all X/Y/Z/rhw values exactly. Six homogeneous D3D clip halfspaces classify 11 birds entirely outside and two intersecting the frustum in every complete retained Worker frame. The analysis has positive, outside, behind-eye and crossing-plane synthetic controls. Counts do not substitute for pixels. The compositor and geometry reads are sequential, not frame aligned; the final geometry repeats the terminal window and is not an independent sample.

The observed API-to-GPU packet path loses no bird geometry in these windows. The current record lacks an independent projection of the original model through the original guest matrices, so it cannot rule out an upstream transformation defect or establish raster/reference parity. The earlier October 31 source/geometry findings in dx-sdk-d3drm-samples.md explain that ordinary time-seeded initialization can place the flock outside the camera; they are historical context, not a current-build oracle.

Source and validation

Runtime pins record Git dfcb4dada4e06862b8d6ed26ae567a77f50cbb50, with relevant-source.patch and untracked-inputs.patch. The live standard module is SHA-256 e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467; the retained compatibility module is 04acbb89f66ad9814836b5e2c34de659b518f4e1994f50c4ee8069c991f02f72. The source pins describe the tested checkout inputs; no fresh compilation or new module-build provenance is claimed. Forty original media/input pins are referenced by existing path and SHA, with no copied media/source archives. No production code changed; no full build was required for this evidence note.

Serial cooperative observation and closeout

After the Worker runtime was closed and its evidence retrieved, a separate run-local cooperative capture path was qualified locally on the real production module and in-page GPU owner. Its absent-app and not-ready controls reject via the same scalar readiness helper; the final 15/30/45-second captures and 17 raw draws/completed submissions pass. An intermediate control accidentally returned the GPU object through CDP and failed with an internal serialization error; returning a scalar fixes that generated-only control. Both failures and the successful final proof are retained. The five supplemental run-specific input files are separately SHA-pinned and verified before the original launch.

The original cooperative scene launched once at 10:54:34.900 UTC, after the Worker parent and Chrome were gone. Its full-compositor 30/60/120-second images were personally inspected: cyan terrain changes, with no visible bird. The three geometry windows retain 6, 7 and 6 complete 67-draw frames respectively; all 13 birds are wholly outside the homogeneous frustum in each. All 130 bird triangles reach completed submissions with exact X/Y/Z/rhw preservation, depth testing/writes enabled and no culling. Both paths report zero GPU errors and zero fallbacks. The in-page context's actual queried identity is WebGL 2, ANGLE Vulkan SwiftShader (Subzero), with a native 640×480 16-bit target. The two runs have different naturally seeded phases; they are not matched-time renderer parity tests.

closed-terminal-retrieval-verification.json verifies 90 remote paths, 74 unique files and 435,344,228 unique bytes, including six original compositor captures, full geometry, logs and both standard/compatibility modules, before the one delegated normal nonforced stop request at 10:57:30.749 UTC. No source or media archive was created. An earlier read-only terminal inspection guarded on a still-open cooperative port just before ordinary closure completed; it failed safely, preserved its outcome and dispatched no input or lifecycle operation. A subsequent distinct inspection after the recorded clean exit confirms all four parent/Chrome PIDs absent, four ports closed, both profiles absent and no profile process, writer lock or pending command. Native input is empty; terminal button/queue/key samples are clear. No unknown input was replayed.

The bounded task establishes ordinary off-camera geometry at the renderer input/submission boundary, including a current terrain-only cooperative scene. It does not reproduce the predecessor's exact time-seeded phase or explain its single blue object. The next discriminator for an upstream transformation question is an original-model/guest-matrix projection oracle on naturally completed geometry. Full flock, lighting/raster parity, FPS, audio, physical GPU performance and historical27 remain unqualified.

The service confirms bx_z65jbe9z archived at 10:57:31.051 UTC, expiry cleared, stop completed and forced=false; custody returns to the coordinator, with no resume requested. Six owned local fixture parent/Chrome pairs, twelve ports, profiles and command/writer directories are also clear. Large geometry JSONs were individually losslessly gzipped after retrieval: each decompressed byte sequence was SHA-verified before unlinking the plain file. This reclaimed 400,353,973 bytes. geometry-lossless-compression.json and the retrieval receipts preserve raw/compressed hashes and sizes; gzip -dc recovers the exact original bytes, and analyze-geometry.py reads the compressed files directly. The source remains Git plus recorded patches. Local free space was 2.84 GB after this owned cleanup.

The global evidence checker retains the pre-existing Alice outside-run error. This run's artifact containment/existence checks pass. Boids is absent from the candidate manifest, so its required result.json is an explicitly unattached named-app investigation with candidateId: null, rather than a candidate card.