Boids WebGL content diagnosis, 17 November 2026
Evidence: scratch/runs/20261117-boids-webgl-paused-content/.
The complete-display capture repair remains published as 53ddc2882. This task
changes only its investigation harness, fixtures and tests. No production
activation, renderer, context, retirement, clock, ownership or capture semantics
changed. No original flock/motion repair is claimed.
The previous observer installed a single global trace at library load. Loading the software executor after WebGL hid the active WebGL executor's trace. It now retains a separate bounded trace per executor and selects the active trace on execution. Tests cover loading both modules, return/exception preservation, copied inputs and bounded storage. A separate passive Presenter wrapper records bounded public frame arguments and completed presentation counts. Diagnostic mode uses existing public API tracing with at most 40 lines per selected name. These changes do not schedule guest work or presentation.
The original Microsoft DX5 SDK self-extractor was retrieved from
https://archive.org/download/idx5sdk/idx5sdk.exe. Its 33,018,416 bytes have SHA-1
b14370372307360a9e8de2ebd8fcd13173fd3b4a, matching the previously documented SDK.
Five relevant original sample sources were extracted, individually hashed, and
retained as reference inputs; no repository snapshot was made. The original
executable remains SHA-256 b9327e6fb2758e232d0ca0fb410b11097fd557dc59b3d405167d976db208b8d9.
winproc.cpp OnPaint calls PaintPaused for any inactive, paused, invalid,
minimized or not-yet-created D3D window. OnIdle separately gates DrawFrame.
Thus the text alone cannot establish current simulation state.
Exactly one passive diagnostic WebGL/actual-Worker original ran on immutable 53ddc288272a34a73fec775c4b419167f2c28e34 plus recorded observer suffixes. Preparation and runtime used the fixed 240/260 and 300/320 second limits and independent prearmed watchdogs. No input, window override, configuration/media edit, forced flush, unpause, private memory access or unchanged retry occurred. No additional software original was needed or launched.
The public trace records natural WM_ACTIVATEAPP(TRUE), WM_ACTIVATE(WA_ACTIVE) and WM_SETFOCUS delivery, followed by continuing BeginScene, EndScene and Flip. The page instance's thread-local focus getter is zero despite delivered focus; that value is not an activation oracle. Three complete captures still show the unchanged Paused display. WebGL receives changing submitted vertex/index bytes, 67 DRAW_UP records in each retained complete PRESENT interval, full-target CLEAR, and alternating render target IDs 3/9, each matching its PRESENT target. Presenter frames reach HWND 65537; delivery counts advance, stale rejection is zero, CPU-present sequence remains zero, and the compositor accepts the current GPU layer. No device/context failure is recorded. Counts and accepted layers are not evidence of visible GPU geometry in this original.
A newly authored public COM Device7 flip-chain fixture runs through the actual guest Worker and WebGL render Worker. Its complete source has exactly the known 320x240 blue geometry rectangle within a 640x480 opaque black frame; all 307,200 source pixels pass the literal oracle. A causal public vertex-color mutation changes that same submitted geometry to black, with matching native worker execution and complete-frame oracle. Nine native ownership/failure controls per fixture reach their required readiness/capture stage and refuse. This proves a working native GPU geometry/Flip route, not sample-specific Device2 parity. The blue fixture's literal pixel oracle was applied offline to its retained capture; the black fixture ran the same oracle inside the harness. All fifteen run-specific harness file versions were reconstructed and hash-verified against their launch pins. Existing compiler normalization changes non-ASCII comment bytes in 88 in-memory WAT buffers; independently reproduced hashes account for those served-source differences from Git. No guest code was patched.
The remaining blocker is localization of the sample-specific pixel loss between GPU execution, transferred bitmap content and final composition. Existing logs prove submission and delivery but do not retain independent pixel observations at those two intermediate boundaries. The bounded two-frame stream trace also evicts initial CREATE/SHADER/UPLOAD definitions, so it is not a self-contained replay. A production fix based only on these records would be speculative. The next useful work is an authored Device2 fixture reproducing the original's 16-bit flip chain, depth attachment, viewport Clear, indexed terrain/bird draws and initial GDI paint, with independent public frame-content checks before and after bitmap transfer. The simpler Device7 fixture does not reproduce the defect. Further original execution needs new authorization or a material causally proved, natively qualified and published repair; the repaired-original authorization was not used. The task stops with this concrete content-localization blocker.
Canonical flock build and focused tests are recorded separately from immutable runtime qualification. The first build log reaches completion but its tool wait returned 143; it is not counted as passing. A subsequent build has an explicit native wait receipt. Runtime/module/media/source hashes, owned binary patches, selected named screenshots and actual process waits are in result.json. No frame series, source archive, media bundle, box, extra worker or public deployment was created. Inherited November 16 joins remain UNKNOWN; browser closure is not guest Quit. Box ID null, state NOT_CREATED, expiry null, stop receipt N/A.