Crusaders of Might and Magic (demo)
3DO, 1999. Fixture: test/binaries/win98-games-a-d/Crusaders MM demo-D3D,
installed tree at installed/ (crusaders demo.exe, gfx_d3d.dll,
gfx_sw.dll, uz.dll = Info-ZIP UnZip32 5.4, Miles mss32.dll;
data in caps\, meshes\, models\, script\, textures\textures.zip).
Image base 0x400000, no relocations needed.
Registered as crusaders_mm_demo: exe crusaders demo.exe, the install
dir's localFileManifest, workingDirectory/exeGuestPath
c:\program files\3do\crusaders of might and magic demo. The manifest
predates a tools/gen-win98-games-a-d-manifests.js entry; adding one needs
a regenerate, which that tool refuses below 2 GiB of free disk.
Route
At boot it asks "Use This DLL?" for gfx_d3d.dll, then gfx_sw.dll
(MessageBoxA MB_YESNO). --input=20000:dlg-cmd:7,40000:dlg-cmd:6 picks the
software renderer. Title (any click) -> main menu (Play Game 132,175) ->
Play Game screen (New Game 158,115). Title and both menus render correctly
in software DirectDraw with its own drawn cursor.
New Game loads the level (~240k batches at the defaults) to the hero in a
stone corridor with HUD health 50 / mana 25. Held arrow keys play: Left turns
him (keydown:37, ~6000 batches, keyup:37), Up walks. Driven as a
frozen --control session with tools/ctl.js (step / click / cmd
keydown / png). Evidence: scratch/runs/20261010T0145Z-crusaders_mm_demo-prototype-w6.
Fixed on the way (commit with this note)
VerQueryValueAreturned every ANSI string in one scratch slot. The game keepsFileVersion, queriesCompanyName, then compares the first against "5.4" (0x4cbf01), so it read "Info-ZIP" and refuseduz.dll("has the wrong version number"). Each string now gets its own slot at half its UTF-16 offset.- A fixed-address
MEM_RESERVEpast the image window (guest0x04000000+with image base0x400000) used to "succeed" with the literal address -- which$g2wtranslates onto emulator tables ($TITLE_TABLEfor0x04000000) -- or fail by accident. It now fails withERROR_INVALID_ADDRESS, as Windows does for a range it cannot place.
The fixed pools, and why the memory map has holes for them
0x450ea0 reserves three pools at fixed addresses: 0x04000000 (512 KB),
0x06000000 (0xEE1000) and 0x08000000 (1 MB), storing the results at
[obj+0/0xc/0x18], then (0x450fc0) reserves AND commits a fourth,
0x29040 bytes at 0x05000000 into [esi+0x24], and exits(-1) quietly
if that one fails. They are not hints. The level files (caps\CaDEMOa.cap,
meshes\CaDEMOa.msh) are memory images whose embedded pointers are already
relocated to those bases: with the pools relocated into the sparse arena, the
command list at 0x58ca48 holds records like 0x0609c0e2 (= pool base
0x06000000 + 0x9c0e2), the texture-register caller at 0x4783ff passes
records with a NULL name 49 times (0x4c34f0 builds "Textures\" + name),
and the interpreter switch around 0x4787ae-0x478bac finally calls
through a NULL vtable at 0x4cb57a (record 0x0800e37d, a pointer into the
0x08000000 pool), batch ~435150 on the route above.
Guest 0x04000000-0x08100000 is inside the direct window, which used to
translate those four ranges onto emulator tables (window tables, the thread
cache, DX/GL tables). The fix is four region.declare-derived
$GUEST_FIXED_POOL_* holes at (g2w 0x04000000/0x05000000/0x06000000/0x08000000)
in src/00-regions.wat, so those addresses are plain guest memory through the
existing one-compare g2w fast path, and VirtualAlloc returns a fixed
request wholly inside a hole ($guest_fixed_pool_of). Room came from pinning
$THREAD_CACHE_BASE at 0x1A000000, in a tail carved from the sparse pool
(316 -> 288 MB for every app). A 42-app coop/threads crash sweep found no
regression: scratch/runs/20261010T0155Z-crusaders-pool-sweep-w6. Rejected:
an app-scoped mutable g2w-fast bound (a global load in every translation for
every app, plus Crusaders' heap/stack on the slow path). None of the pools
overlaps the heap (0x04100000), stack (0x07400000) or thunks (0x07500000).
Unrelated noise seen: \\.\ramlockC.vxd CreateFile fails (it falls back);
loose Textures\*.tga misses are normal, it then reads textures.zip.
Ordinary browser gameplay verified, 2026-11-02
One original invocation of the registered crusaders_mm_demo app reached
controllable corridor gameplay in a visible, headful browser. The ordinary
?candidates&app=crusaders_mm_demo&no-log route used the unchanged installed
fixture, manifest and registered defaults. At the visible native “Use This
DLL?” prompt for gfx_d3d.dll, the focused Yes button was clicked. No
software-DLL prompt followed. Title click -> Play Game -> New Game, retaining
the default Knight difficulty, reached the stone corridor; the settled HUD
read health 50 / mana 25. This run selected the D3D guest DLL, with the
host's default WebGL renderer. The historical CLI software route above is a
separate observation; this run establishes no both-backend parity.
Personally reviewed full-page and present-canvas captures show the hero's heading and corridor view changing during held Left, followed by held Up movement toward the wall. A short Left alignment exposed the corridor's arches, then another held Up moved the hero alongside the wall toward its decorative panel. Before, held, released and later settled captures preserve these changes. Native keydown-to-keyup durations were 977 ms Left, 2275 ms Up, 202 ms Left, and 1753 ms Up (requested 700/2000/200/1500 ms; held-capture overhead accounts for the longer observed holds). All 20 browser input events were trusted, with four balanced click pairs and four balanced key pairs. Consumed key releases followed browser keyup by 10/39/51/15 ms. Final input inspection found an empty queue, no active input, both physical/asynchronous arrow states false, and zero mouse buttons; the later neutral capture retained the final location.
The final camera is very close to the ornamented wall, with an overhang obscuring the upper scene and a leaning/raised-knee pose persisting after release. The bounded result proves menu access, turning and movement in this corridor; it does not establish general rendering/animation correctness, FPS, audio, combat or full-game completion. No guest files, registered defaults, renderer settings, guest state or clocks were altered, and no recovery download or second original invocation was used.
Evidence: scratch/runs/20261102-crusaders-native-browser/, particularly
review.json, retrieved/native-input.json, retrieved/input-state.jsonl,
retrieved/09-corridor-before{,-canvas}.png, the held/released/settled pairs
for 10-left-turn, 11-forward-movement and 13-corridor-forward, and
retrieved/14-final-neutral{,-canvas}.png. All 53 PNGs passed chunk CRC checks;
777 served-file receipts matched immutable pins. The actual gameplay page's
Wasm response was SHA-256
307168f9e0475fd3c10f87cbc5d98f05006c6d4494d2f564381a9a0ebe289fe6
(2,071,475 bytes). All 3,936 accepted locked-build inputs were refreshed before
execution and checked afterward, along with exact source/compiler snapshots,
src/api_table.json, module, installed manifest and all 358 fixture files.
The coordinator's root observed-file hashes establish local identity, not
upstream provenance. This docs-only run reused the matching accepted locked
build; it made no production changes or new build.
Immutable transfer member checks and independent SHA verification of all 123
retrieved evidence files preceded explicitly delegated shutdown of boat
bx_226wj2x7. Its original expiry was 2026-11-02T19:30:06.449Z, recorded
before use; STOPPED was confirmed at 2026-11-02T18:38:31.947Z, with scoped
remote helpers/profile absent. Controller and Chrome exited 0. The local
launch wrapper and original messageboard watcher returned 143 without an
intentional stop or a saved wrapper terminal status; the cause and remote
supervisor's raw exit status remain unknown. This monitoring limitation is
preserved rather than inferred away. The independent live disk guard stayed
running throughout transfer/runtime/retrieval, with no observed floor breach.