Army Men runtime closure follow-up: CPU DLL blocker
Historical source last updated, 2026-10-20. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.
Coordinator-authorized follow-up to 620c56be42ced91642c67b3602cb7929535ec3e0. One original-demo execution on bx_6czhvg6m, 2026-10-20 12:15 UTC. The two previously demonstrated dependency failures are repaired. Actual
Boot Camp paired-scroll acceptance remains unmet: startup calls a null
wincpuid procedure pointer before reaching the menu. No renderer or ordinal
patch, second guest launch, main integration, or worker launch occurred.
Dependency repair and preflight
Reused /home/user/icm-fullvm, retaining its canonical source/module:
Installed pngjs 7.0.0 from the exact repository lockfile URL. Verified
the archive against its lockfile SHA-512 integrity before extraction. It has
no package dependencies. No package/lock or production source edits. The npm archive encodes directories without execute permission (0666);
plain tar extraction initially failed on its own newly created lib directory. The initial error/receipt is preserved. Restored that user-owned directory's
traversal mode and extracted the same verified archive with
--no-overwrite-dir. No access policy or privilege change was involved. Constructed the real raster canvas and Win98Renderer, then encoded a PNG.
Copied these exact historical support DLLs into the documented initial PE
discovery location /home/user/icm-fullvm/test/binaries/dlls/:
Source was the retained original
/home/user/army-men-local-stage-v14-20261019/runtime/test/binaries/dlls/;
each byte matched the prior scroll-entry pins.json. No DLL redownload or
local binary duplicate. An in-memory archive transferred the four files;
remote archive and extracted bytes were verified.
The unchanged resolveDllGraph and detectRequiredDlls, using the runner's
discovery directory order, resolved five PEs (the four above plus the
app-local original smackw32.dll) with zero missing dependencies. This checks
the static import graph, not future names passed to LoadLibraryA. The actual run confirms all five PEs loaded, their DllMain calls returned 1,
and the EXE/MFC imports were patched. The prior MFC42 ordinal trap and
renderer-unavailable message are absent.
One execution and exact next blocker
Used the unchanged reviewed observer preload, cooperative threads,
200,000-block batches, a 750-batch/90-second bound, and prior ordinary
click/Numpad6 schedule. Exact arguments are in recovered/launch.json. The runner exited 0 after 750 batches, but that exit code is not success:
- At batch 9 it logged
[eip-zero] guest called through NULL. dbg_prev_eip=0x004584c8, EAX=0, stack return address0x004584cf; final main EIP remains zero.- The five guest screenshots are identical black 640×480 images.
bootcamp.pngis only the scheduled filename; visual review confirms it does not depict Boot Camp. paired.json:accepted:false, zero records, pinned owner, zero observer errors, zero pending depth. No HRESULT or pre/post scroll samples exist.
Static disassembly and strings from the exact original EXE identify the failing sequence:
458493 push "cpuinf32.dll"
458498 call [66d840] ; LoadLibraryA
45849e mov ebx,eax
4584a0 test ebx,ebx ; nonzero branch reaches lookup
4584ba push "wincpuid"
4584bf mov esi,[66d83c] ; GetProcAddress
4584c5 push ebx
4584c6 call esi
4584c8 mov [5f5380],eax
4584cd call eax ; NULL, return address 4584cf
4584d2 push "cpunormspeed" ; later lookup, not reached here
The 168-file staging manifest omitted windows/system/cpuinf32.dll and
cpuinfo.dll, both present in the historical original installed VFS. The
current LoadLibraryA implementation's absent bare-DLL fallback can return the
EXE image base; the saved stack contains 0x00400000 in the preceding
module-argument position. This is consistent with a synthetic nonzero module
handle followed by a failed GetProcAddress, rather than a missing MFC
ordinal or terrain-copy failure. No new API trace was run; this attribution
combines the saved null-call state, original instructions, static loader
source, and manifest absence.
Retained original bytes were read-only hash-verified locally:
Both originals export wincpuid and cpunormspeed; cpuinf32.dll statically
imports KERNEL32/ADVAPI32, while cpuinfo.dll additionally imports MSVCRT. Their PE reports and original pins are retained. The precise next prerequisite
is restoring the original dynamically loaded CPU DLL closure under the VFS
system path used by findVfsDllBytes, then verifying those named exports. cpuinfo.dll is retained historical companion evidence; this run directly
demonstrates only the cpuinf32.dll lookup. No claim is made that these two
files exhaust all later dynamic dependencies.
Per the bounded task, no additional provisioning or execution followed this new blocker. Terrain stripe causality remains unknown.
Evidence, recovery, and release
Self-contained new evidence:
scratch/runs/20261020-army-runtime-closure-retry/. It includes provisioning/controller JS, dependency-resolution receipt, source
and asset verification, initial extraction failure, recovered run logs,
screenshots, observer report, static PE diagnosis, archive and hash inventory. The recovered archive and all 23 listed artifacts were hash-verified. verify-evidence.js replays the important evidence assertions without a guest. No candidate ID exists in the local corpus manifest for this game, so this
investigation uses receipt.json, not a dashboard record for another game.