Myth: The Fallen Lords (Bungie, 1997) — demo and retail ISO

Status 2026-10-05 (NEW-GAME-MYTH-20261005): both the official Win32 demo and the retail CD reach in-level 3D gameplay headless with ordinary mouse and keyboard input. Evidence: scratch/runs/20261005-myth-tfl-retail-iso-gameplay/ (retail, reviewed, result.json) and scratch/runs/20261005-myth-tfl-demo-gameplay/ (demo; not a registered candidate, so no result.json), and the browser route scratch/runs/20261005-myth-tfl-retail-browser/ (reviewed). FPS not measured.

Registered app myth_tfl (localhost-only)

Build or rebuild the local tree with node tools/prepare-myth-tfl.js (runs the disc's own Setup.exe headless, ~5 min; --manifest-only rewrites just the cue and .wine-assembly-browser.json from an existing installed/). A fresh run reproduces the installed tree byte for byte.

lib/apps.js LOCAL_CANDIDATE_APPS; refused by tools/deploy-berrry.js. Local media under test/binaries/candidates/myth-the-fallen-lords/ (gitignored):

• installed/ — the Small install Setup.exe writes (copied out with --save-vfs-prefix='c:\program files'), mounted at c:\ by .wine-assembly-browser.json (c:\tags\tags.gor, c:\modules\tcpip.dll, …).
• sources/myth-tfl.cue — one MODE1/2048 track over myth-tfl.iso; the cdAudio config mounts it lazily as D: MYTH_TFL (HTTP Range in the page), so the 538 MB ISO is never loaded whole. disc/ (the fetch tool's extraction) is not used by the route.

Browser (headless Chrome, web-input-probe.js --app=myth_tfl, served by tools/dev-server.js for Range support): menu at ~48 s, level ~110 s real time. In real time the enemy is already crossing the bridge by the time the level is on screen. Input verified there: A turns the camera, F8 key list, minimap click moves the view; Enter (Select All) reached the guest as WM_KEYDOWN 0x0D with no visible selection. Unit select/move was verified on the CLI.

Sources

tucows_205995_Myth_-The_Fallen_Lords-Demo on archive.org is the Mac .sit; skip it. Myth II demo (MythIiSoulblighterDemo) was not needed.

Both installers are MindVision Installer VISE (vise32ex.dll); 7-Zip cannot open them, so they must run in the emulator. Demo: 52 MB install, payload is one tags/tags.gor (51,594,033 B). Retail "Small" install is 31 MB (myth_tfl.exe, tags/tags.gor, tags/scrap.gor, DLLs); artsound.gor (320 MB) and cutscene.gor (114 MB) are read from the CD (D:\TAGS\, label MYTH_TFL).

Modules and API profile

• myth_tfl.exe (demo 713,216 B; retail 714,752 B), image base 0x400000. DirectDraw 8bpp software renderer, or its own 3dfx renderer through our built-in glide2x (the default since 2026-10-06; see "3dfx (Glide 2) renderer" below; the disc's real glide 2.4.3 driver is never loaded), DirectSound, DirectInput keyboard device (acquired, state never read in-game — keys arrive as WM_KEYDOWN), WSOCK32.
• uber.dll — Bungie protocol switch (GetIndexedProtocol, Protocol*), LoadLibrary's modules\tcpip.dll. mclient.dll — bungie.net client, imports uber. Both must be real PEs (lib/dll-registry.js APP_LOCAL_DLLS); as stubs the main menu traps on GetIndexedProtocol.
• In-game input: mouse buttons via window messages, sampled once per rendered frame. Headless at 200 ms/batch an in-game frame is ~250 batches (~50 guest s), so a --input press must be held across a frame (mousedown … 400 batches … mouseup); same-batch click and short presses are silently dropped. Menus run fast enough that click works there.

Emulator fixes this title needed

1. _hwrite (api id 4093) — VISE installer writes with it; delegates to _lwrite.
2. CreateFile on an existing directory must fail ERROR_ACCESS_DENIED (5). Myth's dir-exists probe (exe+0x412b40: CreateFile(dir, GENERIC_READ, OPEN_EXISTING); GetLastError()==5||==32 ⇒ exists) got 2, then CreateDirectory → 183 and the tag-file layer reported "error #2 while writing (#0,#1228)", ending in initialize_myth_for_new_map(NULL) failed. (string 0x4a0014) → exit(-1) with no visible message. Retail recursed on C:\TAGS\PRIVATE instead.
3. uber.dll/mclient.dll app-local DLLs (above).
4. Harness: --iso-exe now reports the disc path (D:\Setup.exe) from GetModuleFileName — VISE copies "from its own directory" and failed with "cannot open the source file 'C:\mclient.dll'". --save-vfs-prefix= added: an unfiltered --save-vfs after an ISO install exports every disc file Setup read and filled the host disk.

Named addresses (demo exe)

• 0x411e90 set-error(type, code, fmt, …); 0x411fc0 error-pending check (global [0x4b64c0]), returns 1 ⇒ init fails.
• 0x412b40 directory-exists probe; 0x412750 CreateDirectory wrapper.
• 0x425e60 new-game/map setup; 0x43b570 initialize_myth_for_new_map.
• 0x40d5b0 halt-with-message (prints nowhere visible) → 0x4256f0 → exit(-1).

Reproduction

Retail (after a Small install into retail-installed/, ISO on D:):

node test/run.js --exe=<retail-installed>/myth_tfl.exe \
  --vfs-include='*.dll,tags/**,modules/**,*.ini' --iso=<myth-tfl.iso> \
  --quiet-api --stuck-after=0 --max-seconds=280 --max-batches=84000 --no-close \
  --input=12000:mousedown:320:240,12100:mouseup:320:240,16000:mousedown:320:240,16100:mouseup:320:240,20000:click:404:272,21900:mousemove:310:258,22000:mousedown:310:258,22100:mouseup:310:258,44000:mousedown:320:400,44100:mouseup:320:400,75000:png:/tmp/game.png

Intro ≈ batch 11900, menu ≈ 21800 (after dismissing "no network modules"), briefing ≈ 40000, Crow's Bridge in-level ≈ 75000. F8 = key list; W/S/Z/X move, A/D turn, Q/E orbit, C/V zoom.

Network modules notice (fixed 2026-10-05)

"Networking is unavailable because no network modules were found in the modules folder" came from UBER.DLL's DllMain (0x10002a7a orig): it chdirs to .\Modules, enumerates *.*, LoadLibraryAs each DLL and asks for NMGetModuleInfo, then restores the CWD. Three emulator faults stacked:

1. CLI: run.js mounted the app's files after the static DllMains ran, so C:\Modules did not exist yet (the page loads files first). It now mounts the manifest before loadDlls.
2. VFS: a directory that only exists because a mounted file lives under it was refused by SetCurrentDirectory/GetFileAttributes (_isDirectory).
3. Both hosts: callDllMain runs an initializer synchronously and treated the LoadLibraryA yield (reason 5) as "resume", so LoadLibraryA returned a stale EAX (0x074ff7b0) and the load was serviced only after DllMain had given up. serviceLoadLibraryYieldSync (lib/process-boot.js) now maps the DLL in place when its bytes are resident; anything needing a fetch keeps the async path. (--trace-api still prints the stale EAX for that call: it records the return at yield time.)

After the fix "Multiplayer Game" is enabled in the main menu.

Performance and audio (browser, headless Chrome, 2026-10-05)

• Metric: WinePerf.snapshot().guestFps = PRESENT/s (explicit DirectDraw presents). In-level Crow's Bridge: 29.3–30.2 PRESENT/s across five samples over ~25 s; page 60 fps, ~26–34M blocks/s, throttled 0%. Myth runs its game loop at 30 ticks/s, so this is at its cap. Menu presents only on change (0/s idle). Headless Chrome only: no display/Xvfb on the box, so no headful number; treat as indicative, not as a felt frame rate.
• Audio (fixed 2026-10-06): intro and menu music always played; in-level was silent. Myth mixes into a few looping DirectSound rings through Lock/Unlock. Once its heap had spilled into the sparse VirtualAlloc backing, IDirectSoundBuffer_Lock rebuilt the guest pointer with the direct-window inverse (wa - GUEST_BASE + image_base), which names unmapped memory there: --fault-null counted ~2.8M writes from the mixer (EIPs 0x476e75/0x476fbf/ 0x476e68) to 0xa1b519c-0xa27ccb0, all landing in the NULL sentinel, so every voice refresh played zeros (browser probe: 67 plays/20 s, PCM peak 0, output 0/400). Fixed with $w2g; same window now: PCM peak 0.282, output non-silent 400/400 samples. Regression: test/test-directsound-lock-sparse.js.

3dfx (Glide 2) renderer (2026-10-06)

Myth has its own 3dfx renderer (C:\myth\render\render_3dfx.c) and picks it by itself. 0x419a80 loads glide2x.dll (our built-in module) through 0x419db0, which resolves 33 entry points by name and returns failure on the first NULL (test eax,eax / jnz after each store, e.g. [0x4b63f0] = _grDrawPolygonVertexList@8); it then calls grSstQueryBoards and sets [0x4ab950]=1 ("3dfx available") when a board is reported. 0x46d1f0(type) is "renderer usable": 0 = software always, 1 = 3dfx if available and the memory tier [0x4ac13c] >= 1 (0x42a710: GlobalMemoryStatus.dwAvailPageFile - 8 MB, capped at 32 MB, against tiers at 0x4a0910: 7 MB, then 12 + 1.5 MB). The graphics prefs record (validated by 0x46d2e0) holds the renderer at +0x1c; the defaults routine 0x46d380 takes the highest usable renderer, so a fresh preference file uses 3dfx whenever Glide is complete.

What it needed from us:

• grDrawPolygonVertexList (convex fan from vlist[0], 60-byte GrVertex; Myth fills its list with add eax,0x3c). It is most of Myth's terrain: 2814 calls vs 280 grDrawTriangle per 1000 in-level batches.
• grSstWinOpen(0, 7=640x480, 0, 0, 0, nColBuffers=1, nAuxBuffers=2) (esi=1 is a constant at 0x46027e). The release Glide 2.4 driver on the disc (glide 2.4.3/grtvgr.exe -> Glide/Drivers/Voodoo/Win95/glide2x.dll, _grSstWinOpen@28 at 0x100094d0) only checks nCol+nAux against frame-buffer memory for 800x600/856x480/960x720 on <= 2 MB boards; ours reports 4 MB. We now accept 1-2 colour buffers (1 = single-buffered: front and back are one surface in both backends) and up to 2 aux buffers.
• Sampling TMU RAM that was never downloaded: Myth draws from 0x0..0x13fff before (and without) any download there; the hardware samples whatever the RAM holds, so both backends now do too (zero-initialised) instead of failing.

Evidence: scratch/runs/20261006T003007Z-myth-3dfx-inlevel-final (CLI, --glide-renderer=software); compare the 8bpp software frame in scratch/runs/20261005-myth-tfl-retail-iso-gameplay/05-gameplay.png. Software stays selectable in Myth's own Preferences.

Choosing the renderer (user-facing). A fresh install starts on 3Dfx. To switch, open Preferences from the main menu, pick Graphics → Rendering → Software (or 3Dfx to switch back), and press OK. The next game uses it. Give the menu a few seconds after OK before clicking: Myth reinitialises its renderer there, and an immediate New Game click is ignored. The choice is stored in Myth's own preference file C:\TAGS\LOCAL\PREF\GAME. Whether it survives a page reload depends on the browser keeping the emulated C: drive; that was not tested, because every probe starts a fresh profile.

Browser verification (2026-10-06, headless Chrome, SwiftShader WebGL):

• Default 3Dfx, Threads on: scratch/runs/20261006T010423Z-myth-3dfx-web-threads-fixed.
• Default 3Dfx, Threads off: scratch/runs/20261006T011714Z-myth-3dfx-web-threads-off.
• Preferences shows 3Dfx selected; Software selected + OK, then in-level software rendering (Threads on): scratch/runs/20261006T010821Z-myth-3dfx-web-switch-software.
• test/test-myth-byo-iso-web.js PASS on the 3Dfx default (in-level is the filtered 3Dfx frame; camera turn and audio pass): scratch/runs/20261006T011945Z-myth-3dfx-byo-iso-web.

Threads crash (fixed b7acece2). With Threads on, Glide goes through lib/glide-render-worker.js, which accepted only a 20-byte open packet. 8cd80d5a's 24-byte grSstWinOpen packet (word 5 = nColBuffers) made it throw Invalid Glide open packet, trapping the guest at EIP 0x4602f6 (Myth's grSstWinOpen). Safari reports the page-side frame instead (lib/render-worker.js:54 _receive, which rejects with the worker's error). Regression: test/test-glide-webgl-fallback.js.

Bring-your-own ISO in the page (2026-10-06)

Evidence: scratch/runs/20261006-myth-tfl-byo-iso-run6/ (reviewed). Regression: test/test-myth-byo-iso-web.js (local-only, skips without the ISO; ~4 min). Both were validated on e7c970ef, i.e. on Myth's software renderer: from 8cd80d5a the game picks its 3dfx renderer, and in the browser that is the WebGL Glide backend, which crashed at the time of writing (claude:1863d2b5's open fix) -- rerun this test once it lands.

1. Drop myth-tfl.iso on the desktop (window.wineMedia._fileInput): the insert dialog reads label MYTH_TFL, 538 MB, 312 files and offers 13 programs; the default is the autorun launcher Startup.exe, so pick D:\Setup.exe.
2. "Insert and keep" copies the ISO into OPFS; Setup's C: writes go to that media's OPFS overlay journal ([overlay] checkpoint: persisted N change(s)). The VISE pages advance with Enter; the final "Install DirectX 5.0?" box is Tab, Enter (= No); then Enter for Close.
3. VISE makes its Start Menu group only by Program Manager DDE ([CreateGroup("Myth The Fallen Lords")], [ShowGroup(...,1)], [ReplaceItem(...)], [AddItem(C:\Program Files\Myth_TFL\myth_tfl.exe, Myth The Fallen Lords, , 0, -1, -1, )], with wFmt = CF_TEXT). The virtual Progman used to ACK without acting (and refused CF_TEXT), so no shortcut and no icon; it now writes the .lnk (e7c970ef). The shell then shows a kept icon "Myth The Fallen Lords" (lnk:<media id>:c:\program files\myth_tfl\myth_tfl.exe).
4. That icon launches the installed exe with D: = the ISO (DRIVE_CDROM, MYTH_TFL): intro, menu, New Game, Crow's Bridge, A turns the camera, F8, in-level sound (54 plays/20 s, PCM peak 0.16).

Driver pitfalls (not product bugs): the new icon is appended below an 800 px viewport -- scroll it into view; the desktop's open gesture is two click events within 500 ms (puppeteer clickCount: 2 did not launch it); a Read Me window may be open over the desktop.

Open

• CLI and browser both need a press held across one in-game frame.