Hexen II offline native recovery and gameplay — 2026-11-14
Run: scratch/runs/20261114-hexen2-offline-native-gameplay/.
Exact local native dependency recovery succeeded without network access. The
retained Q3 package.json and package-lock.json matched SHA-256
3ed55cf48ef3a1a00e7971ea528b9b2c991decbbf8917901310d59f3808e936f and
125ea670ef6333b93d4375ed558fc42c24af2da9a4fbe2ed49e929ee519e0976.
npm ci --offline --prefix scratch/runs/20261114-hexen2-offline-native-gameplay/native-deps --no-audit --no-fund
returned 0. All 162 provider files matched the retained Q3 native pins in both
size and SHA-256 before use. NODE_PATH points only to this run's private
provider directory; createRequire from lib/headless-gl.js resolves both
@node-3d/glfw and @node-3d/webgl there. Root dependencies were untouched.
The first native preflight failed its right-edge resize assertion. An unchanged second preflight passed, preserving both outputs. This repeats the Q3 preflight instability; it does not establish a fix. Native renderer reported llvmpipe LLVM 20.1.2 / WebGL 1.0. The emulator's SoftwareGL flag remained 0 in observations.
Under /tmp/wine-heavy-runtime.lock, all 317 accepted emulator source closure
files and module hash matched the accepted SuperTux build. No new full build
was needed. The actual loaded module was observed as SHA-256
dcf6e8f0ca9d361c10fa556fb30835d34cf5bfe4089f84d298cb57fd170dbd42.
Launch HEAD was bcd802d24f7241f5108322b9f1ffefe86457a0cd; relevant dirty hunks
and untracked runtime input are retained as Git patches. All 18 media/manifest/
control pins passed, including registered glh2.exe SHA-256
76d96731e3076441e5c70bce2089cd8f36f25b373214f656aa9d5cd8d532e9dd.
The registered startup registry, media and controls were used unchanged.
The one original run uses --headless-gl --real-ticks, default 1000-block
batches, cooperative scheduling, fixed --max-seconds=1200, and an independent
1220-second watchdog armed before native launch. It waited behind SuperTux and
TIM2. Native menus reached Single Player → New Game → Paladin → Apprentice.
The runtime reported 1141.176 seconds; its actual parent wait completed about
1144 seconds after launch following native Quit; the fixed
1200-second limit was not extended and no second process launch occurred.
Current bindings were read before launch: W/S move, A/D strafe, Mouse1 attacks, Ctrl crouches. Older Ctrl attack observations were not reused as current input. Each decision requires a fresh reviewed full-scene capture, with a one-use SHA-bound review and 30-second admission limit. Finite movement/look/attack commands were recorded before launch; independent UP is armed before DOWN. The controller preserves UNKNOWN receipts and never replays them. Six finite input tests, three actual child/watchdog lifecycle tests, the new fresh/expired/ one-use review test, and current native owner/guard tests passed.
Licence scope remains runtime/API observation and emulator source only: no guest reverse engineering, disassembly, extraction, patch or substitute DLL. No box, download, deployment, other worker or foreign source change occurred. Full27 remains unproved.
Observed result and limits
The first native route reached the courtyard at health 79, then traversed the
ramp, skull door and chamber. chamber-entry-released.png shows two living
spiders at the doorway. The next chamber scene shows health 68; later views
show 37, 3 and native death at 0. No attacks were fired without an identified
visible target. One ordinary Mouse1 respawn returned to the courtyard at 74
within the same process and deadline. This was a native gameplay respawn,
not a second runtime launch.
The second approach reached the chamber at 54. return-pursuers-review.png
shows a foreground spider at right and health 26. A finite relative aim
(+150,+100) and 1000 ms Mouse1 pulse was submitted toward that visible spider;
motion, DOWN and independently prearmed UP all have successful CLI ACKs.
aimed-spider-consequence.png shows player death at 0 and a living spider.
No distinct enemy pain/hit, corpse, kill counter or outgoing damage/kill
consequence was observed. Neither a living attack pose nor living execution
of that final attack is proved: health could have reached zero between the
reviewed target frame and actual input consumption. CLI ACKs alone are not
claimed as guest gameplay success.
The live native Options → Customize Controls screen was inspected read-only
(native-bindings-review.png): Mouse1 attack, W/Up forward, S/Down backward,
Left/Right turn, A/D strafe, C/Ctrl crouch. No assignments or settings were
changed. Some short movement/back-menu pulses had no immediate visible effect,
and some turn corrections overshot. Subsequent scenes were inspected rather
than treating injection ACKs as traversal proof. No generic emulator defect
was isolated or changed.
There were 50 manually chosen decisions, 46 finite pulses and 155 command ACKs, with no terminal UNKNOWN receipt. All decision captures are retained with review descriptions and hashes; they are manually reviewed scenes, not an automated capture series. Review-to-submission age was at most 1.082 seconds. Capture-to-submission age was measured separately: maximum 35.581 seconds for the safe ramp relative-look correction; the admission gate bounded review age, not capture age. This limitation is explicit. Native input consumption counted 44 key DOWNs, 43 key UPs, and two mouse DOWN/UP pairs. Final Y release reached the renderer and was acknowledged, but the guest exited before one last key-UP consumption; no consumed native UP is claimed for that edge.
Native shutdown and retained evidence
Native Escape → Quit → Enter showed PRESS Y TO EXIT
(quit-confirmation-review.png). Y DOWN/UP were acknowledged and the guest
exited 0. Runtime actual wait 0 has separate stdout/stderr EOFs; watchdog
actual wait 0 reports runtime-terminal, with no deadline signal. The detached
supervisor's launch child and the outer supervisor child both have actual
wait 0 and separate EOFs. The heavy lock was released and all five recorded
runtime/watchdog/supervisor/outer PIDs were absent. The detached outer process's
own wait status was not obtained; its absence is not substituted for that
status. The auxiliary messageboard watcher was terminated; its original parent
wait is likewise unavailable.
Postflight rechecked 18 media, 317 accepted source and 162 native provider pins
with zero drift. All 68 loaded accepted source files matched; 36 other observed
loaded files were hashed and rechecked against their current bytes, without
claiming they belong to the accepted source manifest. The actual module hash
matched. Native GL recorded 7475 presents/readbacks and gl_sw_enabled=0;
this is native provider execution using llvmpipe, not a physical-GPU or
performance claim. The first-preflight resize failure remains unresolved.
The result index names 56 selected, manually inspected PNGs. PNG chunk CRC,
inflation and SHA-256 checks passed. Only two redundant passive loading
snapshots were removed; no decision scene was removed. Compact logs, JSON,
source/module pins and Git patches retain recovery evidence. The private
native node_modules is removed after its successful postflight verification;
its exact retained manifests and offline recovery recipe remain. Root
node_modules, media and foreign source WIP were untouched.
Goal status: traversal, two living encounters, incoming damage/death, exact offline provider recovery and native Quit are established. Outgoing enemy damage/kill and living attack execution remain missing. No new box was created: box ID null, state NOT_CREATED, expiry null, stop receipt N/A. No further runtime is authorized by this completed single bounded task. Full27 remains unproved. A separately assigned future encounter attempt would still need visible aim and a distinct outgoing consequence.