Dethkarz demo (Beam Software / Melbourne House, Infogrames 1998, Direct3D)
App id dethkarz_demo, candidate dethkarz-demo (registered 2026-10-11, w5).
Evidence: scratch/runs/20261011T0750Z-dethkarz_demo-race-w5/.
Source and licence
- archive.org item
DETHKARZ,DETHKARZ.zip: 17,402,067 bytes, SHA-1b43457dfe3d18a25822c5048b735d4570ea476ac. - The zip holds an InstallShield 5 setup.
unshieldextracts three groups:- "Dethkarz Main Files":
Dethkarz.exe,IFORCE2.DLL, the.ifrforce profiles andDevice Libraries\; - "Data";
- "GOB File":
Dethkarz.zip, the game archive (GOB).
- "Dethkarz Main Files":
- No licence grant.
City Demo Readme.txtgives only the story and controls. The exe and setup state no redistribution terms. The candidate islocalOnly.
Layout the game needs
Install Path.HKLM\Software\Melbourne House\Dethkarzmust holdInstall Path. The game opens<drive>:\Data\Dethkarz.zipfrom it.- Unset, it fails with
Failed to load GOB @:\Data\Dethkarz.zip. - The registration writes
C:\Dethkarz, and the GOB goes ininstalled/dethkarz/Data.
- Unset, it fails with
- Renderer DLLs. Renderers are found with
FindFirstFileA(".\Device Libraries\*.dll").- The folder name has a space.
unshieldwrites it asDevice_Libraries, so the candidate copies it back asDevice Libraries. - With only
Device_Libraries, the game fails with MessageBox002 - No valid graphics drivers were found. - The folder holds
Vertigo Direct3D.dllandVertigo 3Dfx (Glide 2).dll.
- The folder name has a space.
Startup driver dialog
DialogBoxIndirectParamA(DlgProc0x402be0) offers "3Dfx Voodoo Native Driver" first and "Microsoft Direct3D Hardware Driver" second.- Picking Direct3D shows a "Direct3D Device" combo: Primary Display Driver.
- The mode is 640x480x16.
Glide is the default, and it dies. OK on the default crashes in
_grTexDetailControl@16(unimplemented in09a8h-glide.wat), called with tmu 0, lodBias 4, detailScale 6 and detailMax 1.0.- Detail-texture combine factors are an explicit backend limitation there.
- Pick Direct3D. In the browser, that means choosing it in the dropdown.
Main wndproc
0x49e0c0. Its WM_SIZE arm (around0x49e1d8) maps wParam onto ShowWindow:- SIZE_RESTORED → SW_RESTORE;
- SIZE_MINIMIZED → SW_MINIMIZE;
- SIZE_MAXIMIZED → SW_SHOWMAXIMIZED.
It then calls
InvalidateRectandUpdateWindow.- Before this registration, our ShowWindow(SW_SHOWMAXIMIZED) on an already-maximized main window re-posted WM_MOVE/WM_SIZE (and an extra WM_SHOWWINDOW). That made an endless loop.
- The post queue drains before host input, so the modal dialog never saw
another click. The dropdown opened but no item could be picked
(
before-fix-dropdown-stuck.png). - Fixed in
$handle_ShowWindow: WM_SHOWWINDOW is sent only on a visibility change, and an already-visible zoomed window gets no resize pair.
Headless route (CLI, --batch-size=200000)
- Pick Direct3D with mouse down and up:
- combo arrow at (494,198), batches 300/303;
- "Microsoft Direct3D" item at (300,236), batches 320/323;
- OK at (243,288), batches 340/343.
- Title "PRESS ENTER" by batch 1000. Send Enter (keydown plus
di-keydown:13) at batches 1000, 1150, 1300, 1450, 1600 and 1750:- Game Select → Arcade;
- Car Select (Hyperon Alpha "Arrow"; Begin Race);
- Metro City (Medium) loading;
- grid.
- Camera flyby until "GO" at about batch 3200.
- Up (
di-keydown:38plus keydown) held from before GO does nothing. Release it and press it again after GO (batches 3240/3250).- The car reaches 111 km/h in gear 2 by batch 3400.
- It reaches 219 km/h in gear 3 and position 19/20 by batch 3700.
Route script: route.sh in the evidence folder. With threads off the run is
about 36 s of wall time for 3900 batches.
Phone route (WebKit, iPhone 15 landscape, Threads on)
- The driver dialog is drawn zoomed to the phone width. Real taps work on it: the combo arrow, then "Microsoft Direct3D Hardware Driver", then OK.
- The menus read Enter through
GetDeviceState, not through messages. An overlay Enter whose keydown and keyup came in one JS turn was never seen; 6d595095a holds every overlay key for at least 100 ms. - The overlay's Esc/Enter (top left) sit under the "swipe up here for more screen" gutter until Safari's bars collapse (MOBILE-GUTTER-COVERS-TL-CONTROLS).
- Race: Gas held 113 km/h, stick steers, Brake, Esc pauses. Evidence:
scratch/runs/20261011T1330Z-dethkarz_demo-mobile-webkit-w5.
Native Shift firing response (2026-11-01)
Evidence: scratch/runs/20261101-dethkarz-native-fire-control/ (including
run2/). Visible firing is confirmed with ordinary Shift input in the
original local Direct3D Arcade race. The requested narrower claim of
DirectInput keyboard consumption is not established: the observed Shift
reader is GetAsyncKeyState, while the traced DirectInput state reads are
for the mouse.
- The installed
City Demo Readme.txtsaysFIRE GUN - Shift,CHARGE GUN - Hold Shift, andREAR VIEW - Ctrl. Ctrl was historical speculation, not an alternate fire key. Car Select shows the stock Hyperon Alpha "Arrow", automatic transmission, and Pulse Laser. No pickup, ammo/state write, configuration fabrication, or controller remapping was needed. - Selected Microsoft Direct3D Hardware Driver / Primary Display Driver /
640 x 480 x 16 in the stock dialog (
direct3d-selected.png). Used the registered install registry and original manifest. Four ordinary Enter press/release pairs at batches 1000, 1150, 1300 and 1450 reached Arcade, Course Select, Car Select and Begin Race; the race was active by 3350. No accelerator or steering input was sent in these runs. - First original:
race-before.pngat 3350 has no player muzzle effect and a blue weapon bar. Shift down (keydown:16plusdi-keydown:16) followed by ten batches givesshift-held10.png: a bright white/yellow flash and blue/green bolt immediately ahead of the player car, with the weapon bar emptied. At 3400 the effect has moved down the road; after ordinary Shift release,shift-released16.pngat 3416 has no near-car flash. Other cars and the lap timer continue animating, so frame differences alone are not the evidence. The player-aligned effect and weapon-bar response are. - Second and final original tested the concrete input-path hypothesis:
keyboard firing polls
GetAsyncKeyState, rather than a DirectInput keyboard buffer. The first run's shortGetDeviceStateAPI filter matched no records, so this run used exact API names and paused tracing outside the short trials.run2/control-before.png(3360),di-shift-held10.png(3370), anddi-shift-released20.png(3390) show the same firing effect with state-onlydi-keydown:16/di-keyup:16, and unchanged WM counters. That harness action feeds the shared keyboard polling state; its name is not proof that the guest used DirectInput. - The ordinary combined WM/state trial is paired in
run2/wm-before.png(3590),wm-shift-held10.png(3600),wm-shift-released20.png(3620), andwm-settled60.png(3660). Reviewed images show a new flash/bolt beside the player and a blue weapon-bar reduction, then the bolt farther ahead and finally no visible bolt. The car remains in neutral at 2–3 km/h; no throttle was held. This establishes firing response, not a hit or damage. input-consumption.json,state-*.json, the NDJSON command/response records, and streamedruntime.log.gzpreserve actual consumption. Combined-trial dequeue/route/DispatchMessage counters progress 11/11/8 → 12/12/9 → 13/13/10. Last records are ShiftWM_KEYDOWN/WM_KEYUP(0x100/0x101, wParam0x10) delivered to HWND0x10001; known input queues are empty after release. The first run also records both edges; its one-batch release capture still had WM_KEYUP pending, which was consumed by the later capture.- Actual
GetAsyncKeyState(0x10)calls return to original VA0x4a819d: 0 → 0x8001 → 0x8000 → 0 across the state-only trial, and 0x8001 → 0x8000 → 0 across the combined trial. These are guest API calls and return values, not merely injected or queued events. Twenty tracedIDirectInputDevice_GetDeviceStatecalls instead request 16 bytes at0x074fee3c, returning to0x4a8c22(mouse path). There are no traced keyboardGetDeviceState/GetDeviceDatarecords in those windows, and the renderer's keyboard-buffer client set was absent. Thus the precise limitation on the requested Shift DirectInput proof is the observed original keyboard API path, not a swallowed Shift or absent weapon.
Identity and bounds: predecessor MDK2's fresh terminal artifacts and explicit
09:57:24 UTC runtime/source/build release were verified before the 09:58:14
gate; actual baseline f0eda338f29a84929075fc90d26c4b4465ee2356.
Original Dethkarz.exe SHA-256
2a18a5cff1807b457a6e043e0c4e18368c66a598819927bc6f54ba5e18b8de41,
manifest SHA-256
94b221e38cf7022177ebc514628186462d8fe15baa4e16a091c6d9225f3d248d;
all 15 files in the original candidate tree and 4444 source/build identity
entries matched before/after. Reused accepted module
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4:
4392 accepted inputs and 141 retained actual build reads matched, with no
rebuild or tracked code/helper/configuration change. Each original recorded
128 actual file-read identities, all matching local bytes, including the
EXE, manifest, GOB, Direct3D DLL and sound archive. Discovery also read the
Glide DLL; the selected renderer was Direct3D, not the Glide detail path.
Retained observer helpers passed actual synchronous and asynchronous WASM execution with whole 512 MB shared-memory hashes unchanged, host/native clock identity unchanged, and emitted late event 10001 retained in a 64-event tail. Observation covers synchronous whole-file reads and completed descriptor reads up to 24 MiB; async/stream/mmap and loader-internal reads are not covered. Read records flush at process exit. A WASM compilation/hash observation does not prove compiler source reads. Input-state still reports guest-post-queue and button-count coverage as unknown and has no exact event correlation.
Both originals exited 0 after 192.197 s and 80.305 s; minimum free disk was 2,216,583,168 and 2,159,755,264 bytes respectively. Local native only; box ID/state/expiry N/A. No impact, damage, combat outcome, charged-shot mechanics, FPS, audio, browser/GPU, or broader playability qualification is claimed.
Bounded visible-combat attempt (2026-11-01)
Evidence: scratch/runs/20261101-dethkarz-visible-combat/ (including
run2/). No visible opponent impact, damage or combat outcome was
established in two original-game runs. This is a failed combat attempt,
not another firing-only acceptance. The first gameplay obstacle was retaining
an opponent in the car's firing line while accelerating and steering: reviewed
bolts passed beside the moving targets. A later unexplained driver termination
ended the final run before the next approach could execute.
- Reused the retained native Direct3D Arcade/Hyperon Alpha/Pulse Laser route.
The retained
drive.pytemplate still named Outcast and 100000 blocks; the completed predecessor's actualrun2/node-exit.jsonnamesdethkarz_demo, 200000 blocks and the paused exact-name API filters. Those actual launch arguments were used here, with ordinary WM/state key edges. No tracked helper, source, test, configuration, registration or backend changed. No guest state or weapon/ammo values were written. - First original: acceleration began at batch 3250 after GO. The
aim1-before.png/aim1-held10.png/aim1-released20.png/aim1-settled60.pngsequence (3250/3260/3280/3320) showed acceleration but no identifiable impact. The lateraim2-*sequence (3500/3510/3530/3570) showed a near-car flash, then a bolt farther down the road, and no target reaction. Speed reached 247 km/h at 3870. Ordinary right steering at the first bend failed to retain the racing line: atbend-held70.png(4450) the player car pitched upward against the barrier, andbend-settled40.png(4490) shows it below the track. That movement is not opponent damage. Up was released, then the run exited normally. - The second and final original tested a concrete change: fire earlier at
3350, before the opponents spread, with a small left aim correction.
The pair is
run2/early-before3350.png,early-held-3370.png,early-released3372.png, andearly-after-3430.png. Intermediate captures every 5–10 batches show the projectile separate from the opponent: at 3412 the bolt is approximately x362 and the target x341 near y197. The flash at the player is a muzzle effect; it is not target impact. - After a small right counter-correction, the second shot is paired in
run2/corrected-before3450.png,corrected-held-3470.png,corrected-released3472.png, andcorrected-after-3530.png. The bolt travels to the left of the nearest opponent; at 3502 it is approximately x301/y205 while that car is x382/y206. Intermediate captures through 3512 show no target explosion or damage indication. Position remains 20/20; the player's energy bar remains full, and no opponent-specific damage HUD was identified. Up remains held during these settled fire captures: racing, camera lag and changing perspective continue. Whole-frame changes therefore do not establish hits. - The final successful capture,
run2/pursuit3810.png, shows ordinary driving at 233 km/h. The next requested step to 4010 failed because/proc/3565454/fd/0no longer existed. The driver tool session returned 143; both driver and runtime were absent. The Node exit observer wrote code 0 and flushed loaded-file/module evidence, but the driver'sfinallydid not produceexit.json. Its gzip has no footer or exit tail; the original is preserved alongside a complete gzip of its recoverable prefix andlog-recovery.json. The signal's sender/cause is unproved; no disk or runtime guard error receipt exists. No third original was launched.
Consumption and identity checks: both runs record actual Shift
GetAsyncKeyState(0x10) calls returning to 0x4a819d, with values
0 / 0x8001 / 0x8000 / 0 across press/hold/release, plus consumed Shift
WM edges and released key-state snapshots. Actual DirectInput state reads
remain 16-byte mouse reads (15 in the first traced windows, 43 in the
second); these do not establish DirectInput keyboard-specific Shift
consumption. input-consumption.json retains the actual calls and WM
snapshots. Known queues empty at the captured releases; guest post queue,
button count and exact event correlation remain unknown.
Gate: Incoming 01a196bc-fc1d-7863-8087-8a0327166bb6 fresh launcher
final.txt/result.json and its explicit 11:08:26 UTC runtime/source/build
release preceded the 11:09:08 UTC gate. Requested baseline was 86bf6a3a;
actual main at gate was 09e52ce62b9f55eba6d7b3a182cb2c6bc9a115c2.
The 25-minute deadline was 11:34:08 UTC; runtime ended before it. Local native
box ID/state/expiry N/A; own process shutdown was delegated.
All 4444 source/build identity entries and all 15 original files/manifest
match before/after. Full before/after original archives were member-hash
verified, including the original force profiles; no external control config
was introduced. EXE SHA-256 remains
2a18a5cff1807b457a6e043e0c4e18368c66a598819927bc6f54ba5e18b8de41;
manifest remains
94b221e38cf7022177ebc514628186462d8fe15baa4e16a091c6d9225f3d248d.
Accepted canonical dd43e4a was reused only after all 4392 accepted inputs
and 141 retained actual build reads matched. Each runtime retained 128 actual
file-read identities matching local bytes, including EXE, manifest, GOB and
Direct3D DLL. Discovery also reads the Glide DLL; the selected renderer is
stock Microsoft Direct3D, not the Glide detail path. Actual runtime module
compilation matched the canonical hash; compilation does not prove compiler
source reads.
The exact retained observers passed real synchronous/asynchronous WASM
execution, whole 512 MB memory hashes unchanged, clock identities unchanged,
and late event 10001 retention. Coverage remains synchronous whole-file reads
and completed closed descriptor reads up to 24 MiB; async/stream/mmap and
loader-internal reads are gaps, with exit-only flushing. Manifest check and
scratch evidence verification passed. No rebuild was performed for this
documentation-only result. First original: 198.679 s, exit 0, minimum free
disk 2,512,642,048 bytes. Second startup was 2,512,535,552 bytes; a later
sample was 2,487,459,840 bytes. Both drivers sampled the >2,000,000,000-byte
guard while alive, but the second's minimum receipt was lost with its driver;
no stronger continuous minimum is claimed. No separate startup/RAM gate or
foreign cleanup was used. Own runtime/driver PIDs and groups
3561681/3561680/3565454/3565451 are absent; no watcher was created and no new
wa-child-* temporary files remain. Combat stays open; no FPS, audio,
full-playthrough or broader qualification is claimed. Single assignment stops.
Open
- Glide path:
grTexDetailControland the detail combine factor. - DirectInput keyboard-specific Shift consumption remains unproved; visible
Shift firing and consumed WM/
GetAsyncKeyStatepolling are confirmed above. ole32.dllis reported as "loaded as a real PE but is not on disk" on a fresh checkout. That is harmless here.