Drakan: Order of the Flame demo (Surreal / Psygnosis 1999)

Registry id drakan_demo. Riot Engine, Direct3D 6 (DX6 HAL through D3DIM), DirectInput keyboard and mouse, DirectSound. Fixture: test/binaries/win98-games-a-d/DrakanOrderOfTheFlameDemoD3D (InstallShield 5 Disk1); the installed tree lives beside it in DrakanOrderOfTheFlameDemoD3D-installed/drakan demo (65 MB).

Install (headless)

The Driver recipe (docs/re-notes/driver-demo.md) unchanged:

1. run.js --exe=.../SETUP.EXE --vfs-include='*' --tick-ms-per-batch=5 --capture-launch=cap -> c:\windows\temp\_ins5176._mp.
2. run.js --exe=cap/windows/temp/_ins5176._mp --exe-guest-path='c:\windows\temp\_ins5176._mp' --vfs-tree=cap --tick-ms-per-batch=5 --control=PORT --frozen --save-vfs=inst --save-vfs-prefix='c:\program files': Next (410,394), Next, Next, Next, step ~550k batches of copying, untick README (238,202), Finish.

No emulator fixes were needed for the install or the game.

Running it

• Mount with --vfs-tree, not --vfs-include. The engine builds every path from the directory of drakan.exe (c:\program files\psygnosis\drakan demo\dragon.rrc). --vfs-include mounts companions relative to C:\, so dragon.rrc fails to open and the game sits in its frame loop on a black screen forever -- a plain PeekMessageA loop with no rendering, which reads like an activation or D3D bug and is neither. The registered app gets the right paths from its manifest's vfsPaths.
• First run opens "Riot Engine Options" (a property sheet) and writes HKLM\SOFTWARE\Surreal\Riot Engine\Settings100 (REG_BINARY, 84 bytes: primary display driver, 640x480x16, dithering, shadows, bilinear) on OK. The app's startupRegistry carries that value, so launches skip the sheet.
• A "Drakan Beta Demo Warning" message box comes up first every launch (--input=40:dlg-cmd:1).
• dragon.rfl is the game-logic DLL (base 0x10000000), loaded with LoadLibrary by the engine; dragon.rrc / music.rrc are the resource containers.
• Input is DirectInput: menus need relmousemove + di-mousedown/di-mouseup (window mouse messages do not move its cursor), and play uses di-keydown/di-keyup (37 Left turns, 38 Up flies forward).

Route (CLI, cooperative, default tick): Direct3D intro logos by ~30k-60k batches, main menu by ~90k, New Game is relmousemove:-6:-75 from the menu's starting cursor, Normal is relmousemove:-2:69 from there, level 1 (riding Arokh in the rain, an enemy walks up) ~230k batches later. Evidence: scratch/runs/20261006T1600Z-drakan_demo-gameplay.

Named addresses (drakan.exe, base 0x400000)

• 0x410a97 main loop: PeekMessage keyboard range, mouse range, then any; calls the frame 0x4078b0 every pass.
• 0x4078b0 frame: [0x476fd0] is a state word (0 at WinMain init, 0x407bf0 sets it, shutdown 0x410cf0 clears it); with it 0 the frame only draws a resource image through 0x40f990 when [[0x47b660]+4] is set.
• 0x407bf0 set-state; 0x410cf0 shutdown (set-state 0, PostQuitMessage).

October 7 browser Threads ON/OFF lazy-read correctness acceptance

The owner-side ReadFile-family destination/count fence repair (c87a2ea1, integrated as a1710870) preserves freshly read file bytes when an armed GPU surface later materializes. The earlier page-shadow bypass (d3ec2640) is unchanged. Actual-WASM/VFS before-control failed on overwritten file bytes; the candidate and existing lazy-surface fence regression passed, followed by all production build gates. This does not claim real GPU coverage from the mocked-GL regression.

Both ordinary browser modes subsequently reached the rainy outdoor dragon level with enemies and HUD through beta OK → New Game → Normal. Worker and root personally reviewed these screenshots and matching mode/running state:

• Threads ON: scratch/drakan-host-write-20261007/browser/threads-on-attempt2/story-wait.png. worker:true, threadsWanted:true, running:true. Browser/server closed 08:28:03.697Z, Chrome exited 0, no retained processes.
• Threads OFF: scratch/drakan-host-write-20261007/browser/threads-off-attempt1/story-wait-3.png, captured 08:52:19.802Z. worker:false, threadsWanted:false, running:true. Browser/server closed 08:52:42.892Z; Chrome exited 0 at 08:52:43.137Z; cleanup errors and pending streams were empty and both PIDs were absent.

Both used frozen source c87a2ea192f912f19a7f7114b58d1c37e21e091e and the full-gated private production module 0db725dd0c7cd4a2826159b9802e586f8fe97c70028cc1521416d76e095a2a08 (1,719,323 bytes). Actual served-module hashes match. No canonical build or public deployment was changed by these browser runs. The first ON attempt was a preserved setup-only allowlist failure, not a guest regression.

This completes the narrow both-mode level acceptance for the lazy-read correctness repair. It is not a new title qualification: October 6 already has controlled CLI gameplay. The ON scene later returned to a menu before movement inputs, with cause unknown; OFF stopped after level review. No new movement, FPS, sound-quality or PBO-warning-resolution claim is made. ops/handoffs/drakan-threads-acceptance-20261007.json records source/run receipts and the immutable artifact inventories.