AoE3 setup after native x86 IsWow64Process repair — 2026-11-05

The single genuine XP-profile DXSETUP.exe /silent session advanced past kernel32 initialization and called repaired GetNativeSystemInfo, then stopped at the absent wintrust!WinVerifyTrust export. Actual guest exit was -21; runner exit zero is not installation success. No restart/provider check or original AoE3 session ran. No second setup was attempted.

The predecessor stopped at the missing dynamic kernel32 export IsWow64Process. Its repair is commit 4d1db138fbbdf50d66afe24b5dda41c65762205b, API 4604, explicit ordinal 597. The starting state was that session's actual final target overlay and unchanged 73-entry registry. Existing genuine February 2010 setup media remains at scratch/game-media/aoe3-dx9c-20261105/feb2010-complete/; no media copy, download or extraction is needed.

Evidence is scratch/runs/20261105-aoe3-wow64-setup/. The evidence-only adapter uses the production runner, records the baseline before the first guest batch, retains actual writes through its writable overlay, and suppresses whole-media VFS export. The launch recipe refuses a second setup session. Tracing included both repaired APIs and genuine DSETUP entry reach. The registry was imported once before execution and never restored afterward. Only tiny predecessor-generated logs and their metadata seeded the writable overlay.

Gate, current checks and tested identity

Preparation remained source-only through legacy framebuffer1925 FINAL 20:01:42Z, Conan player1930 FINAL 20:26:55Z and AoW2 callback1945 FINAL 20:34:39Z. Independent release-gate.json passed at 20:34:54Z: 444 recorded PIDs absent, 183 ports refused, 57 profiles absent, four locks free, zero heavy processes and free disk above 5 GB. No shared source/test edit, build, test, browser or guest preceded that gate. Runtime work finished within the 45-minute window.

Current boot/font/native-DLL closure, exact setup imports and generated host import signatures passed. The three setup-module pins still match the predecessor; COMCTL32 ordinal 17 still maps to InitCommonControls. mapped-export-check.json verifies the actual mapped kernel32 x86 JMP: named IsWow64Process equals ordinal 597, and its thunk contains API 4604. The accompanying isolated regression passed native FALSE output, four-byte/canary writes, repeated/unaligned destinations, ABI and process identity/access/lifetime checks. Both repaired handlers exactly match their landed implementations. The required flock /tmp/wine-build.lock bash tools/build.sh then passed.

Setup used main 140a7fdf6e678be61441403be533b6a3a788ec21, with empty relevant tracked runtime diff. Its retained standard module is 2,174,134 bytes, SHA-256 b6c36db2aa538aa8c6e411d1cbe421d7898f791d65003b5ed54b8564495289d2. The retained compatibility module is 2,176,780 bytes, SHA-256 533f3227abc6acbcbc33006bd1ef8e9d8865d45615c451637142be6be9fbcbf0. Exact source SHA, binary diff, relevant untracked-source patch, source hashes, owned new-file patch and actual modules preserve recovery; Git is the source archive. There is no source-tree archive or repeated media.

Actual setup progression and first terminal failure

setup-process.json and setup-terminal.json record PID 3366949 and host interval 20:36:39.885713Z–20:36:42.882513Z. Ordinary license acceptance was already authorized; silent license handling was not visually observed. The genuine DSETUP entry counter at 0x0118806d recorded one hit.

API 2375 resolves GetNativeSystemInfo to 0x0143e4b4; API 2376 resolves IsWow64Process to 0x0143e81c. API 2377 actually calls GetNativeSystemInfo(0x080ffb98), returning to 0x012acb1e. No actual IsWow64Process handler call appears in this setup trace: its repaired export resolution, rather than its invocation by setup, is proved here. The isolated mapped regression separately exercises its implementation.

The new installer-log suffix reports IsIA64(): not IA64, creates C:\WINDOWS\SYSTEM\DirectX, and continues past optional Managed DirectX detection after missing mscoree!GetCORVersion. It finds the existing C:\dxupdate.cab, 95,820 bytes. Its in-place SHA-256 still matches 54d68f154058433865708ee0dbf3ecf2d609ffbd618e84a1056440379494d9fd. That file check does not establish guest trust acceptance.

API 3026 obtains recognized system-library handle 0x01460f4c for wintrust.dll; API 3027 queries WinVerifyTrust and returns zero. The appended error log reports error 2 and Unable to initialize CWinTrust; GetCDXUpdate() then fails. Setup finalizes phase 0 with -21, described by its own log as a cabinet that cannot be trusted. API 4121 executes ExitProcess(0xffffffeb). The verification function was never reached, so this is not evidence of an invalid cabinet signature or an actual trust verdict.

trust-export-probe.json reproduces the missing export in an isolated tiny PE using the exact retained production setup module, without extra WAT exports or a second setup. Normal x86 IAT calls resolve the kernel32 named/ordinal positive control and return a nonzero wintrust library handle, but return zero for WinVerifyTrust, with exact stdcall cleanup. It also proves the generic missing-export path preserves a poisoned last-error value 0x1234; the installer's error 2 is stale diagnostic state, not proof its cabinet is missing.

Actual state, repair boundary and closeout

Only the two inherited log files changed; their old prefixes are verified unchanged and new suffixes are retained separately:

Actual log Final bytes New suffix bytes Final SHA-256
directx.log 4018 1619 7dd070388ed25db8611a7ec2e640a9b575e0a182b6d97921fef2fab20694c793
dxerror.log 2819 1056 117ee53b0f926760ca9d7a952c7849be1f0f666dd6007e2eef7abe4829ac4f59

No file was installed or deleted. Registry baseline/final both contain the same 73 entries, with no pending replacement, registration or restart delta. No restart request was observed. Guest log timestamps use the pinned June 1999 clock; the host interval above identifies this session. I personally reviewed setup-final.png: uniform teal 640×480 desktop, with no completion dialog or gameplay.

The actual overlay additionally journals eight created directory records. The final VFS directory listing retains the DirectX root but omits its temporary dx1.tmp subtree. Both the exact journal and final listing are preserved; journal entries are not treated as proof all temporary directories remain live.

WinVerifyTrust is absent from the current API table and implementation. No native wintrust/crypt32/softpub/mssip32/mssign32 DLL is present in the searched existing test binaries or permitted AoE3 dependency/setup inputs. Microsoft documents WinVerifyTrust as dispatch to a registered trust provider; a successful trust decision is specifically a zero return. A real repair requires provider, object verification and policy behavior, not an export-only success result. Those facilities and their required trust inputs are missing here. A safe causal repair cannot be reduced to a narrow API patch within this bounded task; no production change or trust decision was fabricated. The offline provenance checks from input recovery do not substitute for the guest's trust provider or its policy.

Known static gaps GetPrivateProfileSectionNamesA and SystemTimeToTzSpecificLocalTime, and the preceding continued SFC ordinal 8/9 and SRSetRestorePoint misses, are not presumed fatal. Actual installer logs, API calls and new log suffixes establish the later trust-export boundary. No version/capability/AttemptToRun/installation state, certificate store or protection behavior was changed. No CPU or memory divergence was demonstrated. There was no second setup, restart, provider session or original AoE3 session.

The next useful work is a real policy-preserving WinTrust/provider implementation or genuine compatible provider inputs, before any separately authorized setup attempt. Current installation and gameplay remain unproved.

Final independent cleanup verifies all recorded owned processes absent, control port 19765 refused, four locks free, no profiles created, zero heavy processes and more than 5 GB free. The pre-publication check is retained separately. The evidence containment check passes. AoE3 has no corpus candidate entry; findings remain in this note rather than an unrelated dashboard record. No platform operation was rejected, media copied/downloaded/extracted, public deployment made or laptop Heroes II operation performed. The exact FINAL board entry releases runtime and owned files after publication.

No boxes or subworkers are assigned. Box ID/state/expiry are N/A; coordinator lifecycle ownership is unchanged.