Hexen II browser preflight stopped before original, 2026-11-18
No original Hexen II runtime was launched. Living movement, visible enemy combat consequence and ordinary Quit remain unproved. The required authored dialog-to-GL and cleanup control did not qualify; continuing to the original would have bypassed that prerequisite.
Evidence: scratch/runs/20261118-hexen2-current-browser-live-combat/.
The selected PNGs depict an authored fixture, not Hexen II.
Immutable inputs and bounds
The runtime was served from published Git
18c7f57eb26e3df51a6e8f1e35640cf8d91bf829 using the existing memory-only
Git reader. Shared runtime overlays were excluded. The compiled and retained
module is 2,418,470 bytes, SHA-256
1504acf2ea5cac3ec3b8c627f4edc64c564e62ad37c06928f288bd8ae83b4f37.
This is the compiled/served module identity; an executing Worker bytecode
capture was not reached, so the independently authenticated executing module
identity is unknown. No executing GL backend was qualified.
The registered original's manifest, its fourteen file references, root
installed/glh2.exe, and two registered control files were hashed without
modification (18 references). Original executable SHA-256 remains
76d96731e3076441e5c70bce2089cd8f36f25b373214f656aa9d5cd8d532e9dd.
The current registration seeds HKLM\Software\Hexen2\Flag, unlike the
November 7/12 startup attempts. That registration and the registered modern
controls were preserved; the real MessageBox prerequisite was tested in an
authored fixture instead of removing the original's flag.
The predeclared windows were startup 180 seconds; authored scope 300 seconds; original 1,200 seconds including a 240-second Quit reserve; cleanup 20 seconds; independent watchdog 20 seconds beyond each runtime cap. Capture was bounded at 10 seconds, CDP at 15 seconds, review at 90 seconds, hold at 1,800 ms, independent release fallback at 5 seconds. UNKNOWN closed the scope, with no replay or later endpoint queries. The original window was never opened.
Actual authored outcomes
The unique PE fixture uses the existing authored continuous native GL fixture,
adding a real blocking MessageBoxA before RegisterClass/CreateWindow/GL calls.
It runs through the normal production guest Worker and native presentation.
No runtime source prefix, scheduling/focus/clock change or synthetic success
was used. Browser screenshots are whole headful X11 display captures with
defaultViewport:null.
calibration/startup.pngvisibly shows the native notice and focused OK. One Enter DOWN was sent after whole-frame review. The independent release child armed, then its second PAGE_READ evaluation returned exceptionDetails before UP. It exited 1, with actual wait and both EOFs. The enclosing scope closed immediately and did not request another endpoint or replay Enter. A balanced Enter is not proved. The raw exceptionDetails were not retained; the precise thrown exception is unknown. The initial progress attribution to transient GL readiness was an inference: the reader does contain a GL readiness guard, but that does not establish which exception occurred. The transition and release were not qualified.- A proposed process-only reader failed its offline assertion that the GL
guard had been removed.
calibration2had already been launched by the shell sequence; it was cancelled by filesystem without guest input. This separate setup mistake is retained. The reader was corrected offline, but no closed scope was reused. calibration3/startup.pngandcalibration3/dialog-ready.png, reviewed whole, show teal desktop without a visible notice, 12.058 seconds apart. Their saved before/after observations report the same live process, Worker, dialog window 65537, focus 65539 and publication sequence 5; no GL endpoint exists. Logical visibility is not visible presentation. No input was sent. The scope closed as presentation UNKNOWN, with no further runtime queries. These observations bound the failure but do not isolate a production cause or establish equivalence to the UT idle compositor issue being investigated separately.
The offline helper checks rejected replacement of ten process/document owner fields and passed syntax checks. They do not qualify actual transition, independent UP, backend, ordinary Quit, or gameplay. No runtime repair was authored; no native/browser oldfail-newpass or locked-build pass is claimed.
Cleanup and recovery
Each scope retained actual browser, Xvfb, Openbox, supervisor, launcher, watchdog and enclosing invoker close waits with both output EOFs. All three containment receipts report no leak; profiles and owned display resources were removed. Browser host termination is cleanup, not ordinary guest Quit. The board watcher also has an actual drained wait. No box was created: box ID null, state NOT_CREATED, expiry null, stop receipt null.
Git remains the source archive. Runtime path hashes, relevant binary worktree diff, media references, unique fixture/helpers, retained compiled module, recovery instructions and verified artifact hashes are in the run. The final result records missing executing module/backend identities explicitly. No game media copies, source snapshots, frame series, guest analysis or patches, config/registry/save patches, Wine, Python, agents, boxes or deployment were used. Held BloodII work and foreign UT/AoW2/render overlays were preserved.
Next prerequisite: qualify visible real-dialog Enter through actual GL-owner transition, independent matching release and ordinary Quit in a fresh authored scope. The original gameplay attempt remains unopened.
Raw-pixel correction from the release follow-up
The later run 20261118-hexen2-dialog-release-browser-gameplay decoded the
saved PNGs without contacting any closed endpoint. The two calibration3 whole
PNG hashes differ, but their dialog rectangle (x=65, y=345, w=299, h=125) is
byte-identical to calibration1's visible notice: decoded RGBA SHA-256
00dd825eb41e2087b9ac81374142ac7f021b34a6356567e8e5f67ff815de9e19.
The gray body and black OK glyph pixels are present. The earlier whole-image
transport review omitted those pixels; the claimed absence of the native
notice is withdrawn. This does not establish an invisible runtime or justify
a compositor repair. The missing calibration1 exceptionDetails remain unknown.
See the follow-up note for the independent release repair and fresh original.