Hexen II browser followthrough after owning-Worker MIDI repair, 2026-11-07
The original registered browser route reaches the Paladin world after the MIDI repair. Ordinary forward movement, right turning and Mouse1 melee animation have reviewed visible consequences on the same owning guest/render Workers. No enemy hit, kill, FPS, audible output or parity is qualified.
Evidence: scratch/runs/20261107-hexen2-midi-repaired-gameplay/.
The published owning-Worker MIDI repair abf1b8c5b31c501e81871392256740ffafdcd138
passed its locked full build, actual production Worker lifecycle/integration
suite and focused regressions before this task began. All recorded repair source
SHA-256 pins matched the files staged here. Unlike the preceding browser run,
this original now passes the notice and reaches the reviewed main menu,
Single Player / New Game, Paladin class and Apprentice difficulty. Loading
progresses visibly through natural model meshing and reaches the world.
The original manifest's 14 files, installed root glh2.exe, and manifest itself
match the recorded 16-path inventory. Root executable SHA-256 remains
76d96731e3076441e5c70bce2089cd8f36f25b373214f656aa9d5cd8d532e9dd.
The registered layout deliberately avoids loading the bundled MiniGL beside
GLRoot/glh2.exe; no game files or layout were changed. The investigation uses
runtime/API traces and emulator sources only, respecting the bundled licence.
No guest disassembly, extraction, patches, substituted engine, music disabling,
or cooperative scheduling was used.
Source commit is 173ba1c249a3360d868d24848d6827f68ef3c1e4; relevant binary Git
diff and owned driver additions are recorded as patches. Git remains the source
archive. The 653 runtime/input pins match before original launch. The only
production differences from the preceding browser run are the four MIDI repair
JS producers. Live guest and render Worker modules were recovered through CDP;
both match canonical SHA-256
e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467.
Before allocation, strict actual Chromium controls passed positive numbered request/publish/consume/exact ACK for trusted key, combo, click and drag, balanced release, independent key/mouse release, nine zero-effect identity replacements, stale/slow captures, partial input, UNKNOWN/late ACK sticky closure and wrong ACK rejection. Failure controls assert the specific rejection reason. The complete staged browser driver passed at the final app-selection boundary with zero game launches and the exact proposed 25-minute deadline. All transitive CommonJS imports resolve. A reused command client's unrelated optional CLI OpenGL import was exposed by the closure audit; the existing publisher and ACK implementation were isolated into an owned small helper before qualification/allocation. The unused generated-guest transport import was removed from the scratch driver. No deadline or identity guard was relaxed.
One fresh UUID-idempotent box bx_vcphthe4 was created at 10:21:07 UTC, expiry
11:01:07.426 UTC, with a 40-minute total deadline including retrieval and
explicitly delegated normal nonforced shutdown. There is one original heavy
runtime, launched at 10:21:35 UTC, with native CDP and no viewport emulation or
resize. The browser deadline is 10:45:50.852 UTC.
Actual guest ownership: process 1000, guest Worker endpoint 1 / slot 0,
target 47BB175249B8D95EA165232C7B31173A; owning render Worker target
A175BCDE066B644182C3CFC99B108CD3, GL/WebGL endpoint 1. A passive query of the
live owning Worker's GpuBackend.WebGLBackend instance found WebGL 1.0,
WebGLRenderingContext, ANGLE Intel UHD Graphics 620 / Mesa 23.2.1, context not
lost. This is live provider evidence, not a selector inference. No FPS, audible
output or parity claim follows.
Reviewed gameplay consequences
review/meshing-two.png is the first reviewed world capture, following loading
through models/splinter4.mdl, then many class/weapon models ending visibly near
models/soulsphr.mdl. At 10:26 UTC it shows a stone courtyard, central statue
plinth, ramp at right, armored hands and health 80. Slow loading was allowed to
finish naturally; no time/cache adjustment was made.
walk-forward: ArrowUp held 1000 ms brings the ornate plinth and left stone pillar into the foreground. The player advances across the courtyard; this is a spatial consequence, not an input counter alone. Held and released captures are retained.turn-right: ArrowRight held 600 ms rotates the view from the plinth toward the ramp and carved doorway. The decorated pillar is now to the right. Held and released captures are retained.attack-two: a trusted Mouse1 down/up exactly 1500 ms apart produces a visible windup and spiked-fist extension, then returns to idle hands. Passiveoriginal/attack-observation-two/frame-1.png,frame-2.pngandframe-4.pngwere reviewed. Frame 1 completes during the hold; frame 2 spans release and shows the extended strike. Later frames show idle hands. Health remains 80, and no enemy target, damage or kill was demonstrated.
Both key holds have passive wholly-in-hold samples and independent release.
All five world/action captures retain the exact same process, document, guest
Worker instance/module, render endpoint, render Worker instance/module, canvas
and page renderer identities (gameplay-review.json). The native viewport is
1024×768; the guest changes its framebuffer naturally from 1024×768 notice to
640×480 GL, displayed within 1024×734 beneath the full-page strip. No test
viewport resize or emulation was used. Browser: Chrome 151.0.7922.108, V8 15.1.
Page renderer: Win98Renderer; scheduling remains the original threaded Worker.
Some statue/model areas appear black in these captures. This run establishes
navigation/attack behavior, not rendering correctness or parity; it does not
prove a new emulator semantic defect.
The first balanced Mouse1 command received its exact successful ACK, but its passive recorder expected a native-event hook absent from this driver. The recorder timed out and supplied no attack-animation evidence. Its failure is retained. A distinct second, freshly reviewed ordinary attack used a passive DOM listener without changing input dispatch or guest state. It records trusted events and animation. This was not an UNKNOWN-command replay. The first command's released screenshot alone is not counted as attack proof.
Retrieval and normal shutdown
After gameplay, the final passive check found zero held keys and an empty input queue. Chrome closed normally at 10:29:28.827 UTC, exit 0 without a signal. Terminal verification confirmed all 14 recorded browser/controller/observer PIDs absent, both ports closed, profile removed, writer lock absent, 12 exact command ACKs, and the final observed mouse buttons/queue clear. All 653 staged runtime/input hashes still matched.
All 192 terminal files (13,208,901 bytes), including the unique live module,
were retrieved and SHA-verified before requesting the delegated normal box
stop. Box bx_vcphthe4 is archived, stopped at 10:30:08.758 UTC, operation
completed with forced:false; final expiry is null. No resume or deletion was
performed. Runtime and box shutdown finished within nine minutes of allocation,
well within the 40-minute total deadline. Final custody returns to the
coordinator; there is no running assigned runtime.
No emulator changes were needed. Only this note is committed; evidence, source patches and run-specific helpers remain in the ignored run directory. Next work, if requested, is separately evidenced enemy combat, rendering correctness, audio or performance qualification. Historical October 11 CLI movement remains accepted separately.