Registered Hexen II threaded world and movement, 2026-11-07

The repaired production CLI --threads local-main MIDI route reached a personally reviewed composited courtyard and responded visibly to ordinary forward movement. Black model and HUD regions persist in the captured original. A demonstrated emulator sampling defect is repaired with failing-before/passing-after pixel regression. Its effect in the original is not qualified: the sole authorized original completed before the repair, and was not replayed.

Evidence root: scratch/runs/20261107-graphics-threaded-followthrough/ (ignored). All paths below are relative to this root unless explicitly qualified. No source or media archive was made. Git is the source archive; preserve this run directory.

Exact preflight and original identity

Original staged source: 994008917622964e3ae9768c0bf3e6802607c480 on main, including ancestor eee2adf57 and the identical repaired lib/host-audio.js blob. source-commit.txt, eee2adf57-ancestry.txt, audio-repair-pin.json, runtime-pins.json, empty relevant-source.patch, and the relevant preexisting untracked new-lib_wininit-boot.js.patch record the source closure. The final JS repair is separate in repair-source.patch and repair-tested-pins.json. owned-inputs-prelaunch.patch and owned-inputs-final.patch recover the small run-specific helpers. No modifications to the registered executable were made.

Before allocation, local local-control-exact-threads/terminal.json passed the production CLI/headless-GL/stdin closed-trace composition, including numbered ping, error ACK, snapshot, composited PNG and balanced Enter with independent UP. local-closure-audit.json records 464 files, no missing runtime requires, and explicit inclusion of lib/host-audio.js. midi-local-main.log passes the production local-main mixed-worker callback/lifetime regression and rejection controls; midi-worker-slot0-3.log passes actual Worker slots 0/3 open, callback, header, restart/reset/close and rejection/lifetime controls. The accepted no-input matrix and cooperative original were not repeated.

Sixteen registered media references and SHA-256s are in registry-media-closure.json. Root installed/GLRoot/glh2.exe is 545280 bytes, SHA-256 76d96731e3076441e5c70bce2089cd8f36f25b373214f656aa9d5cd8d532e9dd. The registration feeds the root glh2 layout, excludes MiniGL, and preserves original settings. Existing media were referenced individually, not bundled into an evidence archive. Native dependencies were resolved and hashed before allocation (native-preallocation-verification.json, native-local-pins.json).

Two failed local setup attempts are retained: the first lacked the new observer; the second lacked native resolution. Neither allocated a box or sent original input. Local native packages were then installed from the retained lock and their official install scripts; the exact composed qualification passed.

One fresh UUID-idempotent creation allocated bx_hqm2pn99 at 16:10:57.146Z; initial expiry was 16:35:57.160Z. The independent non-agent normal stop watchdog PID 2387772 was armed at 16:10:56.484Z for 1500 seconds and verified alive before allocation (watchdog-ready.json, watchdog-alive-before-allocation.json). No other original or heavy guest ran.

The remote preflight verified 471 runtime/input pins and 162 native pins. Native 8x8 red readback was [255,0,0,255], GL error zero, destroyed live contexts zero; actual native WGL/headless regression passed without skip (retrieved/context-probe.json, retrieved/native-wgl.json). There is a preserved prelaunch audit defect: comparison of all 162 local/remote native files failed only on generated node_modules/.package-lock.json root-name metadata, and the launch shell continued. The 161 functional native files match exactly; the metadata difference is recorded in native-remote-local-comparison.json. This was not a functional native mismatch, but the original launch was not gated by that comparison's exit status. Future provisioning should make that check fail closed.

Exact original argv is retrieved/actual/threads/launch.json: production test/run.js --app=hexen2_demo --no-build --control-stdin --headless-gl --threads, 600-second cap, batch size 50000, unchanged default clock, trace window 1000000000..15000, scheduler and MIDI/WGL trace flags, separate exit raw PNG and PCM outputs. Passive observation hooks are retained with hashes. There was no forced fallback/capture, guest write, private game handler, clock/config change, or coordinate click.

retrieved/loaded-audit.json matches all 152 observed reads; postflight has zero changes among 471 staged pins, with matching native postflight too. This CLI has no HTTP-served module. The actual read, compiled, and instantiated standard WASM was 2182493 bytes, SHA-256 e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467, 2288 exports. The compile/instantiate observation is in retrieved/gl-summary.json; both WASMs and loaded native .node/.so binaries are recovered under retrieved/. Original native context: 640x480, Mesa llvmpipe LLVM 20.1.2, WebGL 1.0, GLSL 4.50, remote Node v24.22.1. Selection of this provider is not performance evidence.

Personal review and ordinary input

Each capture below was personally inspected before authorizing its next input. The driver writes independent UP before submitting DOWN; no DOWN reply is needed to arm release. control-custody.json, commands.jsonl, original [ctl] output and the *-review.json files preserve ordering and exact replies.

Composited capture under retrieved/actual/threads/ Reviewed content / next action
notice.png GLHexen2 notice with OK; ordinary Enter
post.png MAIN, Single Player selected; ordinary Enter
menu-1.png SINGLE PLAYER, New Game selected; ordinary Enter
menu-2.png Class selection, Paladin; ordinary Enter
menu-3.png Difficulty selection, Apprentice; ordinary Enter
menu-4.png Natural model-meshing/loading screen; passive 20-second wait
menu-5.png First-person courtyard, fists and HUD 80/0; ordinary Up Arrow
menu-6.png Foreground stone wall/plinth now fills view, ramp visible at right; finish
terminal.png Same final position after input release; normal quit

The Up Arrow binding is documented in docs/re-notes/hexen2-demo.md. One 1000ms wall-clock Up pulse visibly advanced from the courtyard toward the stone obstruction; no inferred displacement or speed is reported. Five 150ms Enter pulses and one Up pulse were balanced, all independently armed; 32 numbered guest replies were contiguous and successful. No UNKNOWN, held key, input replay, or second original occurred. Ordinary png: captured the composited screen. final-raw.png is a separately retained 512x256 texture atlas, not a world capture. All ten PNGs decoded and were reviewed (png-decode-review.json).

Scheduler and MIDI observations

retrieved/actual/threads/raw/schedulers.jsonl observes actual backend worker. Its thread arrays are empty at construction, before close and exit. The CLI argv and backend agree; the observer's generic requested:false field reflects an unset manager option, not cooperative selection. No auxiliary guest execution or concurrency is established by these snapshots.

raw/midi.jsonl observes one midi_stream_open, returning handle 786433. MOM_OPEN (967) was dequeued into callback EIP 4464496 with active 0->1, followed by observation of the admitted callback becoming inactive at a different EIP. Thus the repaired local-main open/callback path passed the old guard in this original. Terminal state is nonrunning with no queued headers or pending callbacks. No prepare-header, stream-out, restart or playback call was observed. audio.pcm is empty (zero bytes); this proves neither playback nor audio quality.

Black rendering reproduction and demonstrated repair

The courtyard shows a black statue/model silhouette and black lower HUD blocks; these persist after forward input. This is a visual defect, not visual parity. Bounded passive emulator observation retained 400 draws after world review, texture dimensions/filters/completeness, matrices, primary colors and enables, plus 1,146,884 bytes of unique runtime RGBA texture data. The first 1800 upload records were bounded; the terminal reports 3549 uploads overall. The GL log is losslessly gzipped at raw/gl-state.jsonl.gz (66,335,650 bytes). An observer metadata mistake included typed-array values in those upload arguments, making the bounded log larger than intended. retrieved/gl-summary.json and draw-texture-table.json are compact indexes; they do not replace the raw log.

Texture names 754 (196 draws) and 725 (48 draws) are enabled with no image levels, min filter 9985 (LINEAR_MIPMAP_NEAREST), MODULATE, and positive primary RGB. Nine HUD textures 835..843 likewise have no image levels, with NEAREST and REPLACE. Other sampled textures have defined complete images. This establishes incomplete sampling state and rules out a universally black guest primary color; it does not by itself establish why those texture objects have no defined images at draw time.

Our ordinary shader enabled sampling solely from API enable state. Only the special border shader checked _textureComplete. WebGL's black result for an incomplete texture therefore leaked into desktop GL behavior. Desktop GL skips texture application for a null level-zero image and for incomplete mip chains; see OpenGL 1.1 specification, texture application/completeness and Microsoft glTexParameter remarks.

lib/gl-compat.js now gates both sampling uniforms on completeness and checks for transitions each draw, including binding, image/filter changes and shared storage changes. API enable state is preserved. Existing border behavior remains. The actual WebGL 1/2 pixel regression failed before with black [0,0,0,192] instead of primary [128,64,32,192]; after repair, all 38 environment/completeness cases pass, including null model/HUD states, valid textures, missing/mismatched mips, binding/filter transitions and independent unit 1. The old unit test's expectation that an undefined unit 1 image enabled shader sampling was corrected; its API-enable assertion remains. Border (16 pixel cases), 1D textures (36), fixed-function and multitexture focused tests pass. Locked flock /tmp/wine-build.lock bash tools/build.sh passes (locked-build.log).

This repairs a demonstrated emulator semantic cause of black output under the observed state. It does not restore the absent images or prove the original statue/HUD is visually correct after repair. Emulator WAT upload/storage paths were inspected without guest disassembly; a storage/allocation hypothesis is unproved because the bounded trace lacks the relevant complete upload/error history. No speculative allocator change was made. Remaining boundary: explain missing image definitions, then separately authorize a fresh original to review this repair's visible effect. No FPS, parity or concurrency claim is made.

Recovery, normal stop and publication

retrieval-verification.json records 125 references / 123 unique recovered paths, 77,596,227 bytes; each local size and SHA-256 matched before stop at 16:15:13.696Z. final-retrieval-reread.json rechecks those artifacts. Recover from this evidence directory and its exact Git SHA/patches, not from an assumed box retention period. Preserve the original media paths and hashes rather than copying media into another run. Native lockfiles, binary pins and recovered loaded native modules supply the dependency recovery references.

The guest quit normally with exit 0 at 16:14:13.314Z. retrieved/cleanup.json finds no owned live guest/supervisor process; no browser/profile/control port or remote watcher was created. After verified recovery, explicitly delegated normal stop stop_9796115d9eece096 completed at 16:15:14.213Z, forced=false, error=null. boat-state-fdba1c5f-3f30-4f96-8188-1c12546bd671.json confirms ARCHIVED; current expiry is null and retention is unknown. The independent watchdog then recorded owner-completed shutdown and exited. Local cleanup closes the board watcher and verifies owned processes absent; free /home remains above 2 GB. Publication commits only the three announced code/test files and this note on main, with immediate pull/push. publication.json records the resulting commit.