Boids launch/window/capture harness, 16 November 2026

The exact original browser branch now shares its readiness, identity and complete-display capture driver with an authored native public-COM fixture. This is a harness repair, not a production rendering fix. The November 7 serialization repair and earlier original outcomes remain unchanged.

Evidence: scratch/runs/20261116-boids-native-window-capture-harness-repair/.

Contract

Preparation compiles and freezes Git source and its module in memory under a separate fixed 240-second bound with an independent 260-second watchdog. Only completed preparation permits a new fixed 300-second runtime, with an independent 320-second watchdog armed before browser launch. A preparation failure cannot become an original attempt. Existing deadlines are never extended.

The driver selects the named registered app, waits for its actual visible top-level window and displayed renderer ownership, and pins object identities, process ID, HWND, instance, guest Worker and render endpoints. It refuses absent, foreign, replaced, stopped or ambiguous owners. WebGL additionally requires an initialized actual executor with submitted rendering records or completed draws. A scalar instance/Worker boolean alone cannot admit capture.

The production requestCurrentDisplayCapture returns both complete source and displayed-output RGBA arrays. The driver checks completion, dimensions, owner, readback timestamps and controller delivery under one 5-second capture bound, capped by the original fixed runtime deadline. Failed, pending, UNKNOWN or late results close that driver permanently. No retry occurs.

Native qualification

The authored PE sleeps 1,500 real milliseconds before ordinary CreateWindow, then calls public DirectDraw/Direct3D COM methods. Its ordinary native GetPixel verifies COLORREF 0xcc4433. It then idles using Sleep. The reference expects RGB (51,68,204) within the literal 640×480 native primary, transparent untouched source outside it, and opaque black displayed background. The full output oracle uses the returned public viewport with nearest sampling. Every component of both complete frames is compared, without sampling or changing the reference to match a failure.

final-software-worker passed three complete captures: 1,228,800 source pixels and 1,185,280 output pixels per frame. Native ownership controls reject absent app, wrong process, replaced window, replaced renderer and bad Worker readiness against the real production owners. The readiness record includes polls during the delayed native window creation. Actual page, guest Worker and render Worker modules match SHA-256 83703369cc49c3e16077a4abd43def27262487f83dd52583d1b433adf7095f5f. The output was visually inspected at full size.

The earlier continuous-draw fixture correctly refused an in-flight paint and returned no frame. The initial WebGL fixture timed out at the inherited CDP 12-second bound. That bound was subsequently removed in favor of the fixed runtime deadline. Later passive native-fixture observations found an actual WebGL Worker consuming LAYOUT records, with zero geometry draws; successful native software pixels do not establish native WebGL geometry. Those failed qualification runs remain separate and cannot authorize a WebGL original.

The final capture-webgl-worker subsequently passed three full-frame literal comparisons and all real-owner fault controls. It pins the actual initialized WebGL endpoint and native software window independently. The endpoint reports two LAYOUT records and zero geometry draws. This qualifies the shared capture driver on this requested route, while explicitly providing no native WebGL geometry proof. The new Device7 authored input retains the same literal pixel reference. final-software-coop also passed three complete literal frames and the native owner controls. A final software Worker run pins the identical driver bytes used by the published harness after the WebGL readiness change.

The focused observer/driver controls, complete-display endpoint tests, draw tracing and GPU-stream decoder passed. Canonical fullflock build passed with actual tool exit 0; its shared outputs were released. The immutable runtime uses Git a00eb35bebb621ebb3f2b42b6af18f1bd874b9f6, independent of foreign uncommitted rendering work and shared build outputs. All original media are referenced by existing path and SHA-256, never copied.

Authorized original outcomes

Harness commit 840cd90f72649e7ed1f0c9ecc65253f738a97134 was committed and strictly pulled/pushed before originals. Each configuration launched exactly once with its own preparation and fixed runtime deadlines. All four reached the real selected Boids2 Sample window, HWND 65537, process 1000, 640×480. These are actual per-runtime identities, not guessed constants in the driver.

Original configuration Current result
Software / cooperative Three complete frames, publications 163, 338, 655. Terrain and birds change naturally; the last frame clearly shows a multi-bird colored flock.
WebGL / cooperative First capture refused pending host surface publication. No frame and no retry.
Software / actual guest Worker Live guest endpoint 1 and legacy software render endpoint 1; 67 completed software draw calls at readiness. First capture refused pending host surface publication. No frame and no retry.
WebGL / actual guest Worker Live guest endpoint 1 and d3dim-gps WebGL render endpoint 1. At readiness its executor reports 67 draws, 5,062 vertices, 1,154 triangles and one present. First capture refused pending host surface publication. No frame and no retry.

The selected original images are original-software-coop/frame-0-output.png, frame-1-output.png and frame-2-output.png in that same directory. All were personally inspected at full size. The first shows scattered birds; the second mostly terrain with an edge bird; the third shows the flock. This is bounded original rendering evidence, without a claim that all 13 birds or 130 bird triangles are individually visible or independently identified.

All three complete native sources are 640×480 and displayed outputs 1280×926. Their public viewport records an uncropped 640×480 source at native origin (0,0), displayed in destination (22,0,1235,926). The native window, process and renderer remain pinned across captures. Software/cooperative has no decoded public GPU command trace, so world/view/projection matrices are unavailable for this successful sequence; no private state was read to fill that gap. Failed captures provide no frame-aligned transform proof.

The WebGL Worker statistics distinguish geometry execution from texture upload alone. They still cannot establish visible flock pixels without a complete frame. The launcher forces ANGLE/SwiftShader, /dev/dri is absent, and the advertised Intel/Mesa string is not hardware evidence. The authored WebGL capture fixture itself had zero GPU draws and retained its explicit software-output classification.

All four originals compile immutable Git 840cd90f7 before their runtime budget; their module remains SHA-256 83703369cc49c3e16077a4abd43def27262487f83dd52583d1b433adf7095f5f. All 152 original executable/media posthashes match. The compiler and served source are independent of foreign rendering WIP; no production renderer was changed. The source preparation negative control exits in preparation with no browser or original launch. Two exploratory overstrict geometry-readiness fixtures exhausted their original fixed runtime deadlines; their timeouts remain timeouts. No deadline was extended or reused.

Native parents and independent watchdogs have actual waits and both pipe EOFs. Original Chrome parents closed with code 0. Host browser closure is not ordinary guest Quit. Final PID/port/profile checks, hashes and selected frame/viewport receipts are retained in result.json and its local artifacts. The process audit covers recorded parents/watchdogs/browser PIDs, not every transient Chrome descendant or Worker thread. No box or additional worker was created: box ID null, state NOT_CREATED, expiry null, stop N/A.

The remaining concrete blocker is complete-display capture during pending host surface publication on the three active routes. Its refusal is not an independently demonstrated renderer defect. No original input, camera/config/ media/EXE change, synthetic clock, private game-state probe, diagnostic fence, capture retry or renderer fix was used. Native reference/backend parity, hardware, unrestricted full-flock coverage, full27, gameplay, FPS and audio remain open.