GLTron private headful surface experiment, 2026-11-16

Evidence: scratch/runs/20261116-gltron-private-headful-surface-capture/.

Private headful presentation did not qualify the full-host capture route. The focused native experiment becomes permanent UNKNOWN at the unchanged 3,000 ms observation bound. Browser CopyFromSurface enters only 3,533.453 ms after capture entry, although renderer drawing and Viz surface activation occur before the deadline. No original game, consent click or guest input was sent. The earlier headless/new-surface evidence and every older UNKNOWN endpoint are unchanged. This is an execution-route negative, not an emulator repair.

Route and source

Normal local packages xvfb and x11-apps were installed after refreshing the package index; the first install could not locate them. Xorg/dummy-driver and Openbox were already present. Only newly created private Xvfb displays were used, with MIT-MAGIC-COOKIE authentication, mode-0600 authority files and -nolisten tcp. The final display rejected an actual unauthenticated xdpyinfo connection. No existing X0 session, laptop, box, public service or other user's browser was used. Chrome, HTTP/CDP ports and profiles were independently owned.

The test serves immutable Git 30bb96a5a44b5e67c64b9a65bcc6fdbab49664de directly from memory and compiles that commit's WATX closure in memory. The actual module is 2,279,037 bytes, SHA-256 83703369cc49c3e16077a4abd43def27262487f83dd52583d1b433adf7095f5f. Actual guest/render Worker source and module inventories are retained. Shared build outputs and foreign source WIP are not used by the experiment. The canonical build is a separate required repository check.

The continuous native Win32/OpenGL executable is byte-identical to the earlier authored fixture: SHA-256 a282fde427480359786e060285a52a0cb94517e2df9ab7a40a9b91bb90e43bbf. It retains native message pumping, ChangeDisplaySettings, GL drawing and SwapBuffers, 1,000-step slices and the normal real clock. No idle/finish, renderer warmup, publication change, paint suppression or forced application frame is introduced. The new test derives the fixture/compiler from the released read-only test; its compiler import-hook restoration also removes temporary extension hooks absent before compilation.

Chrome is 151.0.7922.108, binary SHA-256 22e4d5ef0c637a87b59a9baad42625aba3284dc03d4a5908748e253556318c97. No headless switch or CDPScreenshotNewSurface switch is used. The ordinary headful window is sized to the same 1024×757, scale-1 viewport, without viewport emulation or clipping. The private X root is 1280×1024. The final browser launches directly at the full production host, brings that tab forward and confirms actual visible, focused BODY before launching the fixture.

The launcher explicitly selects ANGLE/SwiftShader, and /dev/dri is absent. The reported Intel renderer string does not establish physical GPU use. The actual WebGL backend, native current context and draws/transfers are recorded in native-v6/actual-gl.json: 324 device draws, 162 presents/transfers, zero observed readbacks at the pre-capture inventory. Native GL routed through WebGL and Chrome's software device must not be described as hardware parity.

Observed boundary

Final native-v6 has one native owner (process 1000, HWND 65537), guest endpoint 1, GL render endpoint 1, publication 7 at host inventory, and separate actual guest/render Worker targets. Object/document tokens and recovered modules are in the retained owner, host and Worker records. The consent button is normally visible, enabled and hit-tested in the DOM, at (351.515625,165,131.875,22). No button is hidden, clicked or exempted from an ownership/visibility check.

An independently bounded, ACK-prearmed perf recorder samples only the owned browser/renderer/GPU process set. Ten native entry probes observe capture, old visual request, browser copy, Viz Skia copy, surface result, PNG encode, renderer draw, Viz draw/swap, swap ACK and surface activation. The executable segment's 0x1000 virtual/file offset difference is retained in the symbol/probe records. These are passive probes, not additional GL work or browser commands. Their possible measurement overhead is not removed by this experiment.

Native event Monotonic seconds Since capture
PageHandler::CaptureScreenshot 2876199.916579 0 ms
Old visual-state call 2876199.917024 0.445 ms
Viz swap ACK 2876202.101174 2184.595 ms
Renderer DrawLayers 2876202.102609 2186.030 ms
Viz Surface::ActivateFrame 2876202.106349 2189.770 ms
Viz draw/swap 2876202.106548 2189.969 ms
Browser CopyFromSurface 2876203.450032 3533.453 ms

The exact full-host request is unchanged: Page.captureScreenshot({format:'png',captureBeyondViewport:false}). It is sent at 11:36:19.300 UTC and becomes permanent UNKNOWN at 11:36:22.304 UTC (3,004 ms observed timer delivery, configured bound 3,000 ms). No screenshot reply or pixels are adopted. There are zero later CDP requests, no connection reopening, rebind, replay or input. The late native copy entry is passive evidence after closure. No Viz Skia-copy/surface-result/ PNG entry is observed through recorder completion; cancellation after UNKNOWN limits what that absence can establish.

The activation/draw probes are process-scoped and do not identify a particular surface or fence. They establish that presentation-pipeline work occurred; they do not prove which frame satisfied the screenshot or identify a scheduler defect. They cannot turn the previous headless alternative-branch 1.957 ms browser-copy / 7,075.295 ms further Viz-copy result into a fix. That branch was not re-experimented with here.

Actual full-frame evidence and preparations

XGetImage/xwd captures the entire private root window, including Chrome decorations and the complete host viewport. PNGs are decoded from retained XWD bytes with ImageMagick, without crop or pixel filtering. All four selected full images were personally viewed.

DOM readiness is therefore not proof that those current pixels have reached the X display. No stale root frame is substituted for a fresh full-host CDP image or admitted for input. V2's actual presentation establishes a real headful output path; it does not repair capture or supply a focused-input positive. Its capture also becomes permanent UNKNOWN, with only old visual request entries in the passive trace.

Six bounded native preparations are preserved rather than counted as six qualified comparisons. V1 compiled the correct module but stopped before any browser on a temporary compiler extension-hook cleanup bug. V3 and V5 stopped before capture while sizing the headful window: early outer dimensions were zero or not yet synchronized with the viewport. The final helper sends only one resize from valid visible geometry and waits within the same 3-second preparation bound. V4 had real content focus but its expanded perf recorder missed the unchanged 3-second enable-ACK bound, so no capture was sent. V6 uses a reduced event set with inheritance disabled and obtains ACK before capture. Only V2 and V6 consume screenshot requests. No preparation launches GLTron.

Checks, shutdown and recovery

The canonical flock /tmp/wine-build.lock bash tools/build.sh passed under an observable 180-second timeout, actual tool wait 0. Its log is closed and its outputs are released. The native experiment reads none of those outputs. The focused --verify-negative check passes on V6: correct full request, module/viewport, visible consent, content focus, private-X denial, pipeline events, permanent closure, zero subsequent requests/input, retained full root frames and actual browser/display/probe/child/watchdog waits. Fixture byte equality and JS/Python syntax checks pass. These checks qualify the negative evidence, not native trusted input.

The 500 ms confirmation, independently prearmed 1,750 ms physical UP, 3,000 ms observation and 5-second game/30-second menu freshness limits are unchanged. Actual trusted finite input and owner/focus/occlusion/stale/late/UNKNOWN fault qualification remain incomplete. No overall fault PASS is claimed.

All six native children and independent deadline watchdogs have actual waits and both pipe EOFs. All five owned browsers/displays/window managers and three recorders are joined; recorded profiles, authentication files, display locks, sockets, loopback ports and 702 sampled exact PIDs/TIDs are absent/closed. Subreaper-adopted descendant waits are retained. The inventory does not claim exhaustive coverage before process detachment. All 26 task-created uprobes, including the separate setup probe, were removed. Host shutdown is distinct from native guest Quit. /home remained above 1.5 GB free.

Git and binary/new-file patches are the source archive. Launch-version patches, source/module inventories, actual Worker hashes, media pins, phase evidence, waits, closure audit and selected PNG/XWD files are in the run. No source-tree, media or Wasm archive is added. All 92 existing game media files were hash-verified by path without copying or launching them. Older helper/test code and permanent baseline results remain unchanged.

Conditional original: NOT_LAUNCHED, zero attempts. No current full-host capture/input route is qualified, so there is no cycle identity, turn, natural-round or ordinary native Quit claim. The 1,200/1,220-second original guards and 180-second Quit reserve were never consumed. Full-27, FPS, audio, physical-GPU and parity work remain open. A future investigation must resolve actual presented-frame/visual-state scheduling at the same bounds before trusted-input qualification and any separately authorized original.

Box ID null; state NOT_CREATED; expiry null; stop receipt N/A. This single-task worker publishes its own diagnostic tools/test/note, releases them and stops.