Civilization II: Multiplayer Gold Edition (Win32 retail)

Registry id civ2_mge, tree test/binaries/candidates/civilization-2-mge-win32/ (gitignored): cd/ is the disc, installed/ the game install.

Movies are Indeo 4, and Indeo is a separate install

The advisor, wonder and council movies are cd/Civ2/VIDEO/*.AVI, all IV41 (Indeo 4). Civ2 MGE reaches them through AVIFIL32 + ICLocate/ICDecompress and MCIWndCreateA. It bundles no decoder: the disc has Intel's installer under cd/Win_95nt Indeo/IVI_95NT.EXE, which puts ir41_32.dll (plus ir41_qc.dll, ir41_qcx.dll, ir32_32.dll, iyvu9_32.dll) in c:\windows\system and registers them in SYSTEM.INI:

[drivers32]
VIDC.IV41=ir41_32.dll
VIDC.IV32=ir32_32.dll
VIDC.IV31=ir32_32.dll
VIDC.YVU9=iyvu9_32.dll

Without that install, ICLocate returns 0 and the game skips the movie, just as Windows does on a machine without Indeo. With it, the emulator loads the driver and runs its DriverProc as guest x86 (see "Installable codec drivers" in docs/video-support-design.md). Intel's DLLs are never committed. Take them from the disc by running the installer.

Watch out: the installed files differ in size. ir41_32.dll is 739,328 bytes and ir32_32.dll is 199,168. Swap them and VIDC.IV41 loads the Indeo 3 driver, which opens, accepts the stream and then decodes black frames. Nothing crashes. Check the size first.

Headless Indeo install (2026-09-28)

S = a scratch directory. Three stages, each an ordinary test/run.js run.

1. Self-extractor. IVI_95NT.EXE unpacks an InstallShield 3 setup into c:\windows\temp:

   node test/run.js --exe="test/binaries/candidates/civilization-2-mge-win32/cd/Win_95nt Indeo/IVI_95NT.EXE" \
     --overlay-dir=$S/ivi-ov --screen=800x600 --no-build --quiet-api --no-close \
     --max-batches=3000 --max-seconds=90
   

   Copy the overlay's files out under their real names (setup.exe, setup.ins, setup.pkg, _setup.lib, data.z, _inst32i.ex_, …) into $S/ivi-tmp (the overlay format is index.json + blobs/).

2. setup.exe -sms. It unpacks the real engine and ShellExecutes it. Capture that launch:

   node test/run.js --exe=$S/ivi-tmp/setup.exe --args=-sms --vfs-include='*' \
     --capture-launch=$S/ivi-cap --screen=800x600 --no-build --quiet-api --no-close \
     --batch-size=100000 --max-batches=3000
   

   $S/ivi-cap then holds the VFS tree, including windows/temp/_ins0432._mp, the 32-bit IS3 engine.

3. The engine, with the arguments setup passed it, clicking Next through the wizard:

   node test/run.js --exe=$S/ivi-cap/windows/temp/_ins0432._mp --vfs-tree=$S/ivi-cap \
     '--args=-sms -fC:\SETUP.INS  -z1 -cx -xC:\WINDOWS\TEMP\' '--cwd=C:\' \
     --overlay-dir=$S/ivi-ov-s3 --screen=800x600 --no-build --no-close --max-seconds=100 \
     --input=1400:dlg-cmd:1,1700:mousedown:493:454,1705:mouseup:493:454,2100:mousedown:493:454,2105:mouseup:493:454,2500:mousedown:396:337,2505:mouseup:396:337,2900:mousedown:493:454,2905:mouseup:493:454,3300:mousedown:493:454,3305:mouseup:493:454,3700:mousedown:493:454,3705:mouseup:493:454,4100:mousedown:493:454,4105:mouseup:493:454,4500:mousedown:493:454,4505:mouseup:493:454
   

   (493,454 is Next/Finish on an 800x600 screen. 396,337 is the licence "Yes".) $S/ivi-ov-s3 ends up holding the five DLLs, indeo.hlp, the new SYSTEM.INI and uninst.exe. That overlay, or just SYSTEM.INI plus ir41_32.dll, is the Indeo machine.

Registry quirk. The IS script writes its ICM registration to HKCR\System\CurrentControlSet\control\MediaResources\icm\vidc.IV41. The guest really passes hKey = 0x80000000, so that is the script's own doing and not an emulator bug. It means no Windows finds the codec through the registry, and SYSTEM.INI [drivers32] is the lookup that matters on Win9x.

Proving the decode without Civ2's own route

The simplest player that exercises the path is Half-Life: Uplink's intro, which plays media\intro.avi with MCI play sierravideo wait into a 320x240 window at (160,120) of 640x480. Give it Civ2's ANARCHY0.AVI (480x120, 15 fps, 111 frames) as c:\media\intro.avi in the overlay:

• test/test-icm-indeo4-candidate.js builds that overlay at run time from INDEO_IR41_DLL or test/binaries/candidates/civilization-2-mge-win32/indeo/ir41_32.dll, and SKIPs if either is missing.
• --tick-ms-per-batch=20, capture at batch 900: the stop/close come back 0. The best-matching ffmpeg indeo4 frame (56), point-sampled the way StretchDIBits scales, is 99.98% within 24, max delta 27, and 94.4% within
  8. The rest is YUV→RGB rounding. About 6 s wall.

Getting there took two emulator fixes (details in the design doc):

• ROL/ROR set ZF/SF. Indeo's generated VLC reader loops on ror eax,0x10 / jz.
• $current_thunk_eip was not restored after a nested DriverProc, which made the parked play wait land at address 0.

ir41_32 also calls LocalHandle as the stream ends.

The game's own movies (2026-09-28)

--app=civ2_mge plays opening.avi with nothing extra: no overlay and no --media-mount. test/test-civ2-mge-movie.js pins this. The same route works in the browser with the guest in a Worker. It took three fixes:

• The CD's data track. The registered CUE is mixed-mode, and the movies live on its data track (D:\civ2\video). Registry cdAudio mounts used to mount only the audio tracks. Now both hosts mount a data track as the ISO it is: test/run.js through openParts, and lib/browser-shell.js through HttpRangeProviders sized from the manifest's trackSizes. Both then go through mediaImport.analyzeCueBundle, the same plan a dropped CUE gets. The disc's label is its own (Civ2:MGE v1.0; the Win16 disc's label is blank), not the registry's volumeLabel. test/static-server.js now serves byte ranges for the browser tests.
• Indeo. The registry entry mounts indeo/ir41_32.dll (the file the install recipe above produces) at c:\windows\system, marked optional: true: without it the game still launches, it just has no codec. It also sets HKLM\...\Windows NT\CurrentVersion\Drivers32 vidc.iv41, which $icm_drv_lookup reads when SYSTEM.INI has no [drivers32] line.
• Audio clock. Civ2 paces video off waveOutGetPosition and refills audio only on MM_WOM_DONE to its CALLBACK_WINDOW. The host read the callback record from stale pre-allocator literals (0xD164/0xD16C) instead of $WAVE_OUT_SHARED, so WOM_DONE was never posted and the movie froze after its first 8 buffers (1.49s). Fixed in 4e5bd041, which also keeps a lazily backed AVIFileOpen handle parked, so that closing it no longer cancels the read it is waiting on.

Headless recipe: --batch-size=100000 --tick-ms-per-batch=20. The Diplomatic Heralds prompt (first launch) appears around batch 500. Its OK button is at 403,387 (mousedown at batch 1210, mouseup at 1230). The movie follows; after it comes the main menu with its animated IV41 map.

Not yet done

• Advisor and wonder movies have not been driven. They use the same AVIFIL32 + ICLocate path as opening.avi.
• The Win16 intro's "black" space renders as dark grey (index value 0x1c). That may be the movie's own palette or the 8-bit output palette; it is unchecked.

The Win16 build plays its movies through the disc's 16-bit Indeo (2026-09-28)

The Win16 disc (civ2_win16) ships Video for Windows 1.1 setup media in VFW_INST\. Its IR41.DL_ is Intel's **16-bit** Indeo 4 driver, KWAJ method 3 compressed and with no stored length. SETUP.INF gives the size:

node tools/kwaj.js "…/cd/VFW_INST/IR41.DL_" …/civilization-2-win16/vfw/ir41.dll --size=774960

(sha256 88f156f5…d99721; test/test-kwaj-expand.js checks it.) lib/apps.js mounts it as c:\windows\system\ir41.dll and writes VIDC.IV41=ir41.dll into system.ini [drivers]. The game calls Win16 MSVIDEO ICLocate/ICSendMessage/ ICMessage in src/09e-win16-api.wat. Those load the NE driver and run its DriverProc as a far call.

What the driver needs from the emulator, in the order it hit them:

• LocalInit(sel, 0, cb): its heap goes after the segment's data, and the segment grows. DGROUP is 0x4EEE bytes with a 0xB102 heap. A heap at offset 0 overwrote the data, and DRV_LOAD failed on its own check of [0xa2].
• DPMI int 31h 06h/0Ah/0Bh/0Ch with the G bit. It aliases data segments 11-14 (0Ah) and sets them to limit FFFFFh with G+D, i.e. 4 GB flat. It sets D on code segments 2-9 and B on the stack segment, and reads bases back with 06h.
• USE32 code inside a 16-bit task. The codec lives in segments 2-9 and is entered from seg10:0x17f2 with 66 FF 5E xx (CALL FAR m16:32). The 32-bit CALL/RET/RETF/JMP/ENTER/LEAVE forms are $th_xfer32 (handler 477, 05c).
• ESP as the guest sees it. The register file keeps ESP linear. The seg3:0 thunk switches SS to a flat alias (base 0x02260000 in the run above). From then on, every instruction that exposes ESP's value must use the offset into SS: mov r,esp, push esp, pop esp, lea r,[esp+N] and ALU ops between ESP and another register (xfer32 kinds 13-18). The codec refuses every frame unless lea eax,[esp+838h] / sub eax,esp gives 838h (seg6:0x0ef0). Before that fix ICM_DECOMPRESSEX returned -100 on every frame.
• KERNEL GlobalFix/GlobalUnfix (197/198) and IsBadReadPtr/ IsBadWritePtr (334/335, plus the Huge forms 346/347).
• KERNEL LocalHandle (11), at seg10:0x2618, when the movie stops. Only a skipped or finished intro reaches it; without it, clicking through the intro trapped in the codec.
• AVIFILE AVIStreamFindSample (163), from the game itself (seg66:0x562a), FIND_NEXT|FIND_KEY. It runs only once decoding falls behind the clock. That happens in a browser, and headless at --tick-ms-per-batch=200, but never at the 20 ms recipe above. The Win16 index keeps idx1's AVIIF_KEYFRAME as bit 8 of each record's stream word for it.
• MMSYSTEM timeGetTime (607) must not go through the Win32 handler. The game polls it in a delay loop after the language dialog, the clock-spin park armed, and the Win16 bridge trapped (0xCA16A9F7) when Enter was pressed there.

Where to look when a frame fails: seg10:0x969a maps the codec's internal status (0..0x16) to an ICERR. The 32-bit decode is seg10:0x16a0 → 0x17f2. Segment bases in these runs: seg2 0x021a0000 … seg10 0x02220000, DGROUP (seg19) 0x022b0000.

Headless recipe: node test/run.js --app=civ2_win16 --quiet-api --no-close --batch-size=100000 --tick-ms-per-batch=20 --input=400:mousemove:209:280, 410:mousedown:209:280,430:mouseup:209:280,700:mousemove:322:327, 710:mousedown:322:327,730:mouseup:322:327,950:png:a.png,1200:png:b.png. The intro's starfield, title and galaxy burst decode frame by frame. test/test-win16-use32-stack.js pins the ESP semantics.

Gameplay (city screen, naming)

Route to the first city, controlled: node test/run.js --app=civ2_win16 --no-close --control=PORT --frozen --batch-size=100000 --tick-ms-per-batch=20, then drive it with tools/ctl.js (Enter through setup, then cmd keypress:98 for b on the tutorial's "good site for a city"). ctl key sends keydown/keyup only, so letter commands need cmd keypress:<ascii>.

• USER ShowScrollBar (267), when the city screen opens (its two SB_CTL bars).
• The city screen is a child covering the main window, opened after the map screen's World (454,0) and Status (454,135 178x299) panels. The renderer's deep hit test took the first child in creation order, so every city-screen button click went to the Status panel. The game answers that with "You must close the City Window before the game can proceed". It now takes the topmost sibling by z rank (test-child-from-point-deep-zorder.js).
• City and rename names are read by subclassing an edit and calling CallWindowProc(WM_GETTEXT) with a far buffer (after WM_GETTEXTLENGTH). The packed lParam went to the native edit unconverted, so every city kept the terrain name the buffer already held ("Hills", "Grassland") (test-win16-gettext-lparam.js).

Civilopedia (Win16, 2026-09-29)

The Civilopedia menu does not show a dialog. Civ2 WinExecs PEDIA\GET_INFO.EXE, an 11 MB Authorware 2 runtime (its image is ~750 KB; the rest is an overlay it reads from its own file). The two talk through FindWindow("Get_Info") and ShowWindow, plus a get_info.txt Civ2 writes with the topic. GET_INFO loads PEDIA\CIVJUMP.DLL, a small NE, by full path. It runs as a second Win16 task: see docs/win16-multitask-design.md.

Controlled route: from the setup screens, Enter ×16 (with step 250 between) reaches the map. cmd post-cmd:386 then opens the Civilopedia's Civilization Advances list; the menu item ids are the WM_COMMAND the bar sends. After about 240 batches the list is up. A click on a name (for example 102,117 for Alphabet) opens its page. EXIT on a topic page is at 497,461, and on the list at 425,432.