Aliens versus Predator: Alien demo (Rebellion, 1999)

lib/apps.js id avp_alien_demo (localhost-only), exe test/binaries/win98-games-a-d/Alien vs Predator - Alien demo-D3D/extracted/avp_alien_demo.exe with smackw32.dll. DirectDraw + Direct3D, DirectInput keyboard; runs on the software rasterizer headless.

Getting the files

AvPDemo3.exe is a RAR 2.x self-extractor and the box's 7-Zip has no RAR codec ("Unsupported Method"), so the extractor runs in the emulator:

node test/run.js --exe='<dir>/AvPDemo3.exe' --batch-size=200000 --stuck-after=1000000 \
  --max-batches=100000000 --max-seconds=200 --input=50:keydown:13,53:keyup:13 --save-vfs=<out>

Its dialog defaults to C:\; Enter is OK. Copy every archive member (not windows\ and not the SFX itself) to extracted\; 7z l -slt gives the sizes to check against (30 entries). Before ec58524f this stopped at snd15.ffl: the whole extraction runs inside the OK handler, and the renderer's Enter sent that IDOK as a nested $wnd_send_message, which gives up after 64 rounds ([sync] ABANDONED wndproc ... msg=0x111).

Route

node test/run.js --app=avp_alien_demo --quiet-api --batch-size=50000 --max-batches=4040 \
  --input='1800:keydown:13,1804:keyup:13,4000:keydown:37,4010:keyup:37' --no-close --png=/tmp/avp.png

Logos, then the main menu by ~1700 (Start Alien Demo / Controls / Video / Audio / Exit); Enter loads the colony level (~2200-3300) into first person. Left/Right turn; a 10-batch tap is a large turn.

Walking and the "black view" (resolved: not an emulator bug)

Holding Up from the spawn point turns the 3D view black with the HUD intact. That is the Alien standing nose-to-wall against the big hive-resin trunk straight ahead, which is unlit: hunt vision draws it black. Evidence in scratch/runs/20261006T072800Z-avp-alien-demo-forward-walk:

• --trace-dx-raw now prints each PROCESSVERTICES' depth range. The "black" frame still submits 89 TL vertices with sane depth (sz 0.68-0.99, rhw > 0); the nearest ones (sz 0.68-0.91) are two screen-filling quads the game lit 0xff000000/0xff020100, in front of lit resin at sz 0.97+. The game drew a dark wall, we rasterized it.
• I (NAVIGATION vision) on the same route shows the resin texture filling the screen at point-blank range instead of black.
• DirectInput is clean: the keyboard buffer has only DIK_UP (0xC8) set, the mouse's buffered GetDeviceData returns 0 records, no joystick. The Alien only climbs with Crouch/Climb (RIGHTCTRL) held, which it is not.
• The 16.16 fixed-point idioms the walking-only blocks use (cdq; rol eax,16; mov dx,ax; xor ax,ax; idiv and imul; shrd eax,edx,16) are covered with negative operands in test/test-x86-ops.js.

What made it look broken is the headless clock. The game presents about once every 5 batches, so at the default 200 ms/batch one frame is ~1 s of guest time: a held turn jumps between coarse yaw angles (a mid-turn frame and the frame after release can show the same view) and one forward frame carries the Alien from open floor to the next wall. Walk at tick-ms:10, and switch it some frames before the keypress so the first frame's dt is small too:

node test/run.js --app=avp_alien_demo --quiet-api --batch-size=50000 --max-batches=4461 --no-close \
  --input='1800:keydown:13,1804:keyup:13,3900:tick-ms:10,4000:keydown:100,4050:keyup:100,4060:keydown:38,4460:keyup:38,4100:png:/tmp/walk.png'

At 10 ms/batch Numpad4 turns a full circle in ~200 batches; 50 batches faces the lit corridor, and walking covers it by ~4100 and stops at its end wall.