Age of Wonders II beta demo
Original local package, separate from the already-qualified Age of Wonders I beta demo. October 8 scoped registry, run-result and re-note search found no earlier qualification for this second title. The initial investigation reached the main menu; later sections record accepted gameplay repairs and the November 1 current-main army-control and lower-panel visual check.
Original media and preparation
test/binaries/win98-games-a-d/Age of Wonders2 demo-SW.exe:
101,537,511 bytes, SHA-256
1244f0114965d011d1e28b97e207db15c902d124ebb25af8a6c97748beb73dc0.
The collection's existing provenance is in test/binaries/SOURCES.md.
No alternative edition or additional download was used.
This executable contains a normal ZIP catalog: 1,104 entries, 1,063 files,
197,002,538 uncompressed bytes. tools/extract-age-of-wonders2-demo.js checks
the original identity, catalog totals, names/collisions, each ZIP CRC and the
game hash, and refuses an existing output or crossing the 2 GiB disk floor.
The remote copy was named /tmp/aow2-original.exe; the transfer receipt binds
that temporary name to the original package above.
The output is extracted, not an asserted installed state:
Age of Wonders2 demo-SW/extracted/. All files remain byte-identical. The
package's aow2Log.txt is a developer's old Windows 2000/Radeon log, not our
runtime evidence. The original aow2Setup.exe remains in the payload but was
not executed; no settings or Windows registry state were fabricated.
AoW2.exe: 5,748,736 bytes, SHA-256
a10590e5dbd013d154b00ea53e66670f4e74d38ab33adb2523f0a662af8f89f7.
PE string FileVersion/Product fixed version: 0.92.0.1837 (fixed file
version 0.92.1.0; string ProductVersion 1.0.0.0). Configuration executable SHA:
bcb06f08e6e5d4e0b41a93b9d77de107e72494a382ebf46ad65bc6f95ab52bbc.
Registered local-only app: age_of_wonders2_demo, guest executable
c:\aow2demo\AoW2.exe, working directory c:\aow2demo. Explicit imported
Borland packages: vcl50.bpl, vclx50.bpl, Ml42ND50.bpl. The generated
manifest contains 1,059 companion files with sizes and lazy loading. Reproduce:
node tools/extract-age-of-wonders2-demo.js
node tools/gen-win98-games-a-d-manifests.js --only=age_of_wonders2_demo
node tools/gen-win98-games-a-d-manifests.js --only=age_of_wonders2_demo --check
The full durable local tree's 1,063 hashes were compared with the remote extraction receipt. No proprietary payload is committed or publicly deployed.
Generic startup repair
First ordinary registered browser launch trapped on SysReAllocStringLen
at runtime EIP 0x00abc57c, before the menu. Baseline loaded module SHA-256:
d8d4096f957ed51cecab589b5a1ec7bf402336f13959ce0710faf2f7d3c68960.
The new handler allocates a counted UTF-16 BSTR, copies before freeing its predecessor (including self-aliasing slices), writes the byte-length prefix and terminator, and updates the caller's pointer only on success. NULL source allocates uninitialized contents; NULL old BSTR is supported. Overflow is rejected before size arithmetic; insufficient allocation leaves the old pointer/content intact. NULL pointer-to-BSTR remains an invalid call. The API is append-only ID 4475 and OLEAUT32 ordinal 5 in both resolvers. Contract: Microsoft SysReAllocStringLen documentation.
test/test-bstr-reallocation.js drives the real resolved thunk/Win32 dispatch.
The corrected baseline harness fails with RuntimeError: unreachable; the
fixed harness passes aliased slices, embedded NUL/surrogate UTF-16, terminator,
NULL source/old BSTR, zero length, overflow preservation and stdcall cleanup.
The first harness attempt used a wrong import-signature filename; that failure
is retained and is not the meaningful baseline. Root subsequently requested a
valid-length actual heap-OOM case (cch=0x3FFFFFF5 in 512 MiB). Its assertions
were added after the fixed native deadline; the expanded test was not run.
Do not treat overflow coverage as allocation-failure branch coverage.
Canonical build gates passed, including the API/dispatch freshness, handler ESP, silent-stub, logical-operand and WATX/data-segment checks. Initial build closure omitted two checked-in ToyVM bundles; they were supplied inside the original build deadline. No optimization, silent-success stub or guest-state patch was introduced.
Reviewed browser findings and limit
Boat bx_997uef9b, Chrome 151.0.7922.108, Node v24.18.1, 512 MiB guest memory,
software renderer requested and GPU flag false. Source base
0d781689e8782fd6d0bce43e9021c749e6a30577, exact source/file receipts preserved.
Fixed served/loaded module:
e4d59e88c62dd17cafaa836120e35e891cd56b497bd80190c980696adc98396c.
- Registered
?app=age_of_wonders2_demo&debug&no-threads&d3d-renderer=softwarereaches a rendered splash. Main thread remains at VCL50 runtime0x00ac63b4, yield 7 (empty message wait). Three cooperative guest threads remain active and a timer is pending. Ordinary splash click reaches HWND0x10002; Enter/Escape events were logged with HWND 0. These inputs did not expose a menu. This does not establish a causal thread/timer bug. - Same browser navigated to the ordinary registered
&threadsroute, without resetting its guard. Four Worker guest threads reach the visible main menu (Campaign, Scenario, Load Game, Replay Intro, Quit). The readme prescribes Scenario → Single for the included scenario, Inioch's Legacy. - Trusted click at viewport
(486,517)on Scenario briefly removes buttons; a later reviewed capture returns to the main menu, rather than exposing Single or scenario setup. A separate ordinary 180 ms down/up produces a captured redraw. No later capture exists for that held click, so its eventual outcome is unknown. No gameplay scene, army selection or displacement was observed. Do not claim input bypass, playability or new-game credit.
Worker-route snapshots retain hidden main form HWND 0x10003 and visible child
0x10004, with the fullscreen/splash HWND 0x10002 as the main window. This
is a possible input/window ownership lead, not a diagnosed cause. The
remote closure lacked test/binaries/tlbs/stdole2.tlb, registered
binaries/dlls/oleaut32.dll, comctl32.dll, and dynamic olepro32.dll.
Built-in APIs were used; those omissions remain limitations, not proven causes
of the Scenario behavior. No alternative fixture/stub was installed to hide them.
VCL50 mapping in the cooperative capture: runtime 0x00abb000, original image
base 0x40000000. Thus 0x00ac63b4 corresponds to original 0x4000b3b4;
baseline failing 0x00abc57c corresponds to 0x4000157c. VCLX50 maps at
0x00dad000 from 0x402f0000; ML42ND50 maps at 0x00ef0000 from 0x00400000.
Worker-route bases differ; its log records thread entry points separately.
Immutable bounds declared before launch: transfer 600 s, extraction 600 s, native build 600 s, installer 1,200 s, gameplay 1,200 s, retrieval/cleanup 180 s, aggregate 4,800 s. Installer phase unused. Native deadline 15:04:37.274 UTC was preserved. Browser guard 15:15:38.887 UTC was preserved across repair and navigation; terminal cleanup completed 15:15:39.001 (114 ms closure overhead). First browser active 41.459 s, second 680.946 s; no overlapping browsers. Transfer 313.438 s, final 4,329,217-byte evidence retrieval/cleanup 5.486 s. All eight owned process IDs and Chrome were absent, exact baseline listeners restored, and 3,569 runtime source pins verified before owned prefixes/archives were removed. Root retains boat lifecycle ownership; expiry 16:47:45.876 UTC.
Sealed self-contained captures:
scratch/runs/20261008T1518Z-age-of-wonders2-demo-before/scratch/runs/20261008T1518Z-age-of-wonders2-demo-after/
Next review: run the expanded heap-OOM regression, then investigate original Scenario input/ownership with complete declared system support and ordinary input in a separately authorized bounded run. No further runtime in this worker budget. Gameplay, sound, FPS, physical presentation and release are unqualified.
2026-10-08 sole-worker native review and menu-route continuation
The expanded actual heap-OOM BSTR regression now passes, including valid
cch=0x3FFFFFF5 pointer/content preservation, on exact source 1455cc2f4.
Sealed receipt: scratch/runs/20261008T1546Z-age-of-wonders2-demo-native-review/.
This completes the previously unexecuted native review; root separately reviewed
and integrated core/registration. No game rerun was used for the native stage.
A distinct bounded registered default-Worker investigation passed the canonical
build with both checked-in ToyVM bundles and served the same e4d59e88… module.
It did not reproduce the predecessor's reviewed menu: reviewed images show title
background without buttons, pre-input startup stack trap 0x074ffd4e, or worker
exit. No Scenario input or gameplay was achieved. The predecessor held-click
outcome remains unknown. Including declared stdole2 and then temporarily omitting
it for a causal A/B did not establish a prerequisite; it was restored. An initial
observer SharedArrayBuffer decoding error invalidates that attempt, and capped
logs limit interpretation. No runtime fix was justified.
VCL TThreadWindow HWND 0x10007, message 0x8fff, dispatches its synchronized
method from lParam+0x20/self+0x24. Two read-only observed dispatch requests overlap
an active slice; arbitration is an unproven source lead. Investigate with a
controlled source regression and actual Worker CPU/stack capture before another
ordinary-input gameplay attempt. Browser shadow CPU getters are not reliable.
Full limitations, source identities, commands, reviewed images and cleanup:
ops/handoffs/age-of-wonders2-menu-route-20261008.md and sealed
scratch/runs/20261008T1623Z-age-of-wonders2-demo-menu-route-investigation/.
The one browser closed at its immutable 1500-second guard; evidence was retrieved
and hashed before owned-prefix cleanup. Gameplay remains unqualified.
2026-10-09: callback ownership fixed; Start faults in Miles stream callback
Main28a104092 prevents ordinary Worker slices from consuming a suspended SendMessage callback's EIP-zero sentinel. Actual command evidence and a failing control/passing real-Worker regression established that defect. Ordinary Scenario -> Single now reaches Inioch's Legacy setup. Campaign is explicitly unsupported by the demo. Start still fails; no player-controlled world yet.
The browser's ExitProcess C000DE05 text was misleading: SEH termination calls
the same host exit import with access violation C0000005 OR DE00. The passive
owning-instance observer in run20261009T0229Z-age-of-wonders2-owning-fault
captured the first CPU fault before termination, with no observer errors:
- Win32 thread2, EIP C53D4000 (unmapped), ESP7EC5FF98, EBP0, ESI016C350C.
- First stack word0118900E. Runtime console pins Mss32.dll at01168000; original PE base21100000, so the return is original2112100E.
- Original21121008 is
call dword ptr [esi+0xe0]; it pushes that exact return. This identifies the stream callback path, with field address016C35EC. The target follows from the instruction/stack; the complete object was not dumped. _AIL_register_stream_callback@8: original2110B6B0/runtime011736B0. Its internal setter211223F0 writes the supplied pointer at21122421 to[esi+e0].- Open/close/service exports are pinned in the run's
mss-export-map.json, alongside the original DLL and relevant disassembly.
Next establish the registered pointer and last writer/lifetime of that object before the indirect call. Do not suppress the callback or disable audio as a compatibility fix. Observer reads384bytes, frame unavailable because EBP0, one original fault then SEH termination; later null-instance file-open errors are secondary. Browser39438 exited0 and boat bx_va3qxqm3 stopped02:36:59.165Z. No gameplay/FPS/audio qualification. Handoff contains full source/run identities.
Stream lifetime capture sites
Static review of the same original Mss32 SHA d0db426ebac97dc410b5130c9c74d19bc04151236df85b012823d6cb3d7fb87c:
- Open wrapper calls internal21120A40. It allocates0x114 bytes at21120A59, then zeroes all0x45 dwords at21120A80. Capture returned EAX at21120A5E and ESI after initialization; do not assume a stable stream allocation address.
- Close wrapper calls21121620. It unlinks the stream through next+104, coordinates service counters, releases buffers, then zeroes the complete object at211217C4 and passes it to the free wrapper211029D0.
- Besides the observed direct call21121008, trampoline21122430 obtains a
stream through sample user data and tail-jumps
[eax+E0]at21122444. - A search for every literal+E0 store also finds sample-related code in 211115xx–211118xx. Equal member offsets do not establish equal object types; these are not evidence that those stores corrupted the observed stream.
Private capture preparation: scratch/aow2-stream-lifetime-20261009/observer.js.
It uses existing owning-Worker block tracing across the stream routines,
records registration arguments/current callback/full0x114-byte object at
selected block entries, and preserves the real log import. Limits:300seconds,
256records and256KiB reads per Worker. It checks the known setter signature
before trusting the prior load base. Pure-JS contract verifies argument and
callback capture, unchanged memory, import forwarding and trace cleanup.
No runtime validation yet. Tracing changes dispatch/timing, and block entries
are not every instruction: missed sites and timing sensitivity must be reported.
This can distinguish registered-invalid versus later-invalid pointers or close
before use; it does not by itself identify every possible last writer.
Actual lifetime trace: valid registration, later bulk overwrite
Run 20261009T0248Z-age-of-wonders2-stream-lifetime reproduced ordinary
Scenario/Single/Start with the private observer. Twelve selected-site records
show stream016C350C allocated on thread1 with callback0, then registered with
0046CB08. It is closed with that valid callback, allocated again at the same
address, and registered with0046CB08 again. Thread2 subsequently reaches
callback-status/before-callback with C53D4000 in the callback field. No close
site was observed between the second registration and this invocation.
The full0x114-byte capture is decisive: corruption extends well beyond E0. At invocation, repeated32-byte groups contain float-looking coordinate values and repeated grayscale-looking words00727272/00737373/00777777. This suggests vertex-buffer data, not a malformed registration argument, but the responsible writer and allocation/backing relationship are not established. Some stream fields are still present, possibly rewritten by service code after corruption. Next capture writes/buffer ownership covering016C350C through016C361F; do not assume graphics causality or patch the callback value. Check logical guest allocation overlap and guest-to-WASM backing alias as separate hypotheses.
All46 artifacts were hashed and reread. Prepare37222 exited0, browser38193 exited0 at02:50:28.300Z; capture used original files and unchanged runtime plus the documented diagnostic overlay. Tracing can change timing and observes block entries only. No gameplay/FPS/audio qualification.
Writer identified: undersized D3DFVF_LVERTEX backing
Run 20261009T0257Z-age-of-wonders2-stream-writer adds a private bounded
watch log at block boundaries. Main-thread change from0046CB08 toC53D4000
occurs after original guest block4432A2, before4432FD. Other threads report
the same value while waiting: those are observers, not the writer.
Original4432C4 writes [edx+8]; captured EDX016C35E4 addresses the watched
callback016C35EC exactly. Stack locals pin destination base016C2384, returned
Lock size1180(4480), source countA0(160), destination stride32, source stride24.
The copy writes5120 bytes into a4480-byte buffer and reaches the neighboring
stream at016C350C. This is guest-address overlap, not merely similar bytes.
Caller442CAC obtains this buffer via vertex-buffer Lock. Its creation branch
442C34 requests FVF1E2, the legacy D3DFVF_LVERTEX format, which includes
D3DFVF_RESERVED1(0x20). Our d3dim_fvf_stride omits that reserved DWORD,
calculating28 instead of32; allocation and Lock size follow that calculation.
The primary contract is Wine's d3dtypes.h,
whose LVERTEX definition includes XYZ, RESERVED1, DIFFUSE, SPECULAR and TEX1.
Next fix allocation stride and packing/unpacking offsets for the reserved DWORD. Add a control-failing actual VB Create/Lock regression with160 vertices and neighboring-allocation integrity, plus color/UV conversion checks. Then rerun ordinary Start without diagnostic WAT/Worker changes. No fix tested yet. Fifty hashes verified; original AoW2.exe and writer/caller disassembly sealed. Prepare37471 exited0; browser38085 exited0 at02:59:15.875Z. Boat bx_2sdgzsm6 is retained for the next isolated native/build phase, expiry03:54:26.401Z; do not start another browser concurrently or confuse retained box with live test.
RESERVED1 fix: ordinary army movement verified
The generic FVF fix adds the reserved DWORD to storage size and skips/writes it at the corresponding packing/unpacking offsets. The real VB Create/Lock test fails on control (4480 versus5120 bytes) and passes with the fix for both VB and VB7, including160-vertex uploads, neighboring-allocation integrity, two-vertex color/UV packing, and unpack bounds. Canonical build, VB allocation failure, ProcessVertices, v3 VB drawing and D3D7 lighting suites pass. The ProcessVertices test now resolves dispatch IDs from api_table.json rather than overflowing the production string pool with test-only names.
Run 20261009T0306Z-age-of-wonders2-gameplay uses no diagnostic WAT/Worker.
Ordinary Scenario -> Single -> Start reaches the world. Dismiss Julia's panel,
select the party at the tower, click adjacent terrain to plot a path, then
click again: the army moves and movement points fall20/20 ->13/20. Reviewed
city-click.png and army-moved.png record the before/after. This qualifies
initial player-controlled gameplay, not release readiness: terrain overlaps
the lower UI and black polygons remain; FPS/audio are not measured/qualified.
All67 artifact hashes reread; module bb9d6e9621a06efb5498e2903fc0a3185967ac76cd33fc76b96c11d714c0a387. Browser41887 exited0 at03:10:42.630Z; boat bx_2sdgzsm6 stopped03:11:23.355Z.
Terrain over the lower UI: fixed by viewport clipping (2026-10-10)
Cause. The game draws its lower UI panel into the back buffer once, then redraws only the map viewport each frame. Each frame is BeginScene/EndScene, followed by CPU Lock/Unlock and Blt UI updates and two BltFast presents. Direct3D never writes outside the device viewport, but both D3DIM rasterizers clipped only to the whole render target. The terrain mesh therefore painted over the panel, and its off-map parts left black wedges.
Fix. D3DIM draws now clip to the viewport (9a438f1b9): in the software
spans and rects, and through the WebGL scissor via d3dim_gpu_describe
fields 34-38. With the fix, the browser shows a clean panel on both
renderers, and the army moves (13/20); see
scratch/runs/20261010T1145-aow2-viewport-clip-fix.
Ruled out. Lazy sync makes no difference with it really off
(20261010T1045-aow2-overlap-ab). Note that the ?no-lazy-sync URL
parameter is not applied at load: it is parsed inside setD3DRenderer(),
where q is undefined. To turn lazy sync off, set
window.WINE_D3DIM_LAZY_SYNC=false before launch.
Browser route (tools/web-input-probe.js --threads on
dev-server.js --isolate, guest canvas pixels):
wait:70000.- Main menu at 800x600: Scenario (569,346), then Single (290,78), and wait 20 s.
- Setup screen at 1024x768: Start (896,697), and wait 45 s for the world.
- Julia's panel: (518,473). Select the army at (515,344). Destination (632,388), clicked twice.
The fixture is local-only (191 MB). boat-ship-tree.js (in the run folder)
ships it as parallel base64 parts in about 2 minutes.
The cooperative CLI reaches the world (fixed 2026-10-10, 918f25772).
- The device lifecycle. Startup creates the Direct3D device through
IDirect3D7_CreateDevice(..., &obj+0xC4)(return address0x444752, VMT0x439df4slot 0xBC). The 1024x768 switch runs insideTThread.Synchronize: T3 sends CM_EXECPROC to TThreadWindow 0x10007, and vcl500x4003060eis the return address of that send. That method releases the device, sets the mode, builds the new primary and back buffers, callsWaitForVerticalBlank(BLOCKBEGIN)(the last call of the mode routine0x53a2e0), and only then re-creates the device. - The bug. The cooperative scheduler's
_dispatchCooperativeSendabandoned the dispatch at the vblank park (yield 13). The device stayed NULL, and Start called[obj+0xC4]->SetRenderTargetat0x445314through NULL; main died at EIP 0, and T3's next Synchronize waited forever. - The fix. The vblank park is now continued like a render wait.
test/test-cross-thread-send-vblank.js; evidence inscratch/runs/20261010T1530-aow2-coop-start-fix. - Telling the two sides apart. Startup shows two
WaitForVerticalBlankcalls, the park and its re-entry, and thenCreateDevice. The broken switch showed one, followed byPeekMessageA.
Open, unrelated to the overlap.
--screen=1024x768drops the Start click. The input router still has 0x10002 at 800x600 while the png is scaled.?no-threadsin the browser does not reach the menu within 70 s.- Browser
--trace-apiin Threads mode used to log all-zero arguments; fixed in 6d4dfdaa1 (the Worker sends its call-time stack words).
Current-main ordinary CLI visual evidence (2026-11-01)
Run scratch/runs/20261101-aow2-current-original-control/ closes the scoped
AOW2-RENDERING-COVERAGE-20261009 current-main gap for ordinary army control
and lower-panel terrain clipping. Actual tested main was
9b70a7d5bf70bb5e395f51e023d9f7862af9d6d2. Battlezone's fresh terminal files
and explicit runtime/source/build release preceded all runtime/build/source
access. No historical movement, viewport or vblank repair was reimplemented.
The original registered age_of_wonders2_demo ran locally with --no-build --control-stdin --frozen, default registration/backend/cooperative scheduling,
100,000-block batches, and no screen override. The accepted build was reused:
all 4,392 accepted input pins and 141 prior observed reads matched; actual
compiled module SHA-256 was
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4.
All 4,444 source/module pins and the full 1,063-file original extracted tree
plus its manifest (1,064 pins) matched before/after. The EXE retains the
5,748,736-byte / a10590e5…af8f89f7 identity above; no compilation archive was
used as a substitute for companion-tree identity.
Two originals, both exit 0, are retained. The first completed batch 2,000, then an overlarge 5,000-step request outlasted the local control waiter. An orderly SIGTERM after 197 seconds pending ended it at actual batch 3,292 with an 800×600 menu capture. Thus this was a control-wait limitation, not an established game startup fault. The second used smaller requests and completed Scenario → Single → Start → dismiss Julia → select army → click destination twice. No duplicate game inputs were sent after a waiter timeout. The later 500-step loading request also outlasted its waiter but completed at batch 4,036 with Julia visible; the queued capture recovered normally. One control shell later ended 143 while waiting for ten steps. The isolated runtime survived; its completion receipt was retained in the raw log, and the capture was recovered without repeating the destination input.
Control PNGs are the actual 640×480 composited canvas, scaled from the
800×600 menu / 1024×768 setup and world. Screenshot-based canvas inputs were
Scenario (456,277), Single (233,62), Start (560,436), Julia (324,296),
army (322,215), and destination (395,243) twice. Actual routed message
coordinates confirm Start (896,697), army (515,344), destination (632,388).
Fourteen scheduled button events were dispatched through DispatchMessageA;
the final trace has 21 dequeues/routes and empty scheduled/pending/cross-thread/
renderer queues. Unknown button/post-queue counters are not evidence of
emptiness. No custom window or guest-memory probes were used.
Reviewed run2/before.png (4,048), selected.png (4,060),
destination-plotted.png (4,072), after-second-click.png (4,084),
arrived.png (4,104), and settled.png (4,124) show the party selected at the
tower, a plotted path, then the party displaced to the right. The selected
party movement display falls 20/20 → 5/20; the final unscaled 1024×768
run2/final.png confirms 5/20. The party marker moves approximately
(332,183) → (420,208) in the composited captures, relative to the fixed tower.
From plotted-before-move through after/arrived/settled, two left mountain
patches (2,800 and 3,150 pixels) are pixel-identical. Earlier illumination
changes and newly revealed shroud are visible, so whole-scene equality is not
claimed. Both lower border patches (4,025 / 3,197 pixels) and the left label
area (4,375 pixels) remain exact across all six captures. Direct review shows
the terrain ending above the panel, with no terrain paint across its controls;
this is evidence for this scene, not every map or rendering path.
The read-only identity observer passed actual synchronous/asynchronous WASM
execution, whole 512 MB memory equality, native/import clock identity, and
emitted late-10,001 retention checks. Exit-flushed observed path reads numbered
553 / 572 (431 / 450 original-tree reads), all freshly matching their bytes.
Hooks cover successful string-path readFileSync, callback readFile, and
promise readFile calls; descriptor/stream/internal-loader reads outside those
hooks remain unobserved. The first run lasted 222.920 seconds, the second
808.974; minimum observed free space was 2,245,582,848 bytes. Own runtime,
driver, groups, control waiters and temporary child directories were checked
absent. No box was allocated (ID/state/expiry N/A). Evidence includes commands,
compressed raw logs, input receipts, images, pins, review, cleanup and checks.
No browser, GPU, FPS, audio, full27 or broader playability qualification follows.