AoE3 native system information setup progression — 2026-11-05

The one genuine XP-profile February 2010 DXSETUP.exe /silent session resolved the repaired GetNativeSystemInfo export, then failed on the next missing export, kernel32!IsWow64Process. That API is now implemented and tested against the emulator's native x86 process model. Setup was not retried. Installation, restart/provider verification and original AoE3 gameplay remain unproved.

The previous session entered genuine DSETUP and failed at the absent dynamic export. Its actual final overlay and unchanged 73-entry registry are the starting state, with no installed DirectX state inferred. The existing 146-file setup remains mounted at scratch/game-media/aoe3-dx9c-20261105/feb2010-complete/; no media copy, download or extraction is authorized or needed.

Evidence is scratch/runs/20261105-aoe3-native-system-info-setup/ (ignored). No media copy, extraction, download, source archive, box, subworker, public deployment, protection change or laptop Heroes II operation occurred.

Gate and tested setup identity

Preparation stayed source-only until Conan residency1825 FINAL 19:17:29Z, Aranna loader1840 FINAL 19:27:34Z and AoW2 browser1850 FINAL 19:35:35Z. Independent release-gate.json passed at 19:36:10Z: 394 recorded PIDs absent, 157 ports refused, 44 profiles absent, four locks free, no heavy processes and free disk above 5 GB. No shared source/test edit, build, test, browser or guest execution preceded that gate. The 45-minute execution limit began there.

Current boot/font/native-DLL closure and exact import checks passed before the full build under flock /tmp/wine-build.lock. The three small setup modules match their existing pins; COMCTL32 ordinal 17 still resolves to InitCommonControls. native-export-check.json proves API 4603 and explicit kernel32 ordinal 596 exist and the handler exactly matches repair da0919a1d1b9452ab91e31b43771f5c03a1acb09.

Setup used main eff3a6aa25d17b2f0666af4cc5cfa8dbdf0061f1, with an empty relevant tracked runtime diff. The standard module was 2173782 bytes, SHA-256 78a9e8d7563c94b1e002a688c5e65d4269946ab8dd9eb68aa3edf271f4cb1c24. Both shipped binaries, exact source SHA, relevant binary diff, source hashes and the small relevant untracked-source patch are retained. Git is the source archive. The genuine redist pin remains f6d191e89a963d7cca34f169d30f49eab99c1ed3bb92da73ec43617caaa1e93f; current trust/revocation remain unknown.

Actual progression and first failure

The single-session recipe adapts the predecessor's evidence-only adapter around the production runner. It retains actual writes in an owned target overlay, records baseline/final inventories and registry, and suppresses whole-media VFS export. It refuses a second setup launch. The genuine DSETUP entry counter and API tracing observe execution without changing guest state. Ordinary license acceptance was already authorized; silent license handling was not visually observed. The actual 73-entry registry was imported once before execution, never restored afterward. Only inherited tiny logs and their metadata were seeded into the new target overlay; base media remained in place.

setup-process.json and setup-terminal.json record the exact production command, PID 3299214, and host interval 19:37:22.850961Z–19:37:23.797293Z. Both IsNTAdmin calls returned true; the genuine DSETUP entry counter at 0x0118806d recorded one hit. Actual installer logs show SetupForDirectX() and its Installing on WinXP SP2 branch.

The first missing lookups remain SFC ordinal 8, ordinal 9 and SRSetRestorePoint, API calls 2013–2015. Setup logs LoadSfcDLL() failure and continues. API 2375 then resolves GetNativeSystemInfo to the actual kernel32 mapped export 0x0143e4b4. API 2376 immediately queries IsWow64Process and returns zero. CKernel32 initialization fails before calling the resolved GetNativeSystemInfo function; this run proves export resolution, not a genuine-setup invocation of that handler.

The newly appended error log identifies IsWow64Process, error 127, followed by IsIA64(): Unable to initialize CKernel32. Setup finalizes phase 0 with value -9, internal/unsupported error. Actual guest ExitProcess(0xfffffff7) is -9; host runner exit zero is not installation success.

Only the two inherited logs changed; no file was deleted or installed:

Actual log Final bytes Appended bytes Final SHA-256
directx.log 2399 1043 f19418fcfc60b21aa1da6f09a3366cf0a3a27f2e3f035eeb88b8d6fcf586449a
dxerror.log 1763 647 8735dee2f2e1e8b8fbfc9869f4e9b4faf7061d9e3ef5bf64313d129477f78f1c

Each old log prefix is verified unchanged and each new suffix is retained separately, preventing the inherited previous failure from being mistaken for this run's failure. Guest timestamps use the pinned June 1999 clock; the host interval above identifies the actual session. Registry baseline and final both have 73 identical entries. There is no pending replacement, registration or restart delta. No restart, provider check or original AoE3 session ran because installation failed. I personally inspected setup-final.png: a uniform teal 640×480 desktop, without a completion dialog or gameplay.

Causal API repair and validation

Microsoft documents IsWow64Process as a two-argument kernel32 BOOL function: successful queries return nonzero and write FALSE when the process runs under 32-bit Windows. XP requires PROCESS_QUERY_INFORMATION access. This runtime models native x86 without a WOW64 layer; the JavaScript/Linux host architecture does not determine the guest answer. Its handler writes a four-byte FALSE output and returns success only after process and output validation. It preserves last error on success.

Current pseudo/durable handles use the existing current-process validator; child/alias identity, access and closed lifetime use the existing host process authority with access bit 0x400. Failed access/handle queries return zero, preserve the output and propagate the authority's error. Unwritable output ranges use the existing page-access probe and return ERROR_NOACCESS. The existing model does not track rights/closed lifetime for OpenProcess's current process durable handle; this repair does not change that authority or claim broader access accounting. The native x86 result does not synthesize DirectX installation, version, capability or AttemptToRun state.

API ID 4604 is appended and explicit kernel32 named ordinal 597 is added without renumbering old ordinals. The API hash, dispatcher and mapped kernel32 image are regenerated with existing tools. All six production paths and the new regression are explicitly owned.

test/test-is-wow64-process.js first failed at the absent dynamic export, reproducing the genuine setup failure through the real PE loader. After the repair it passes in standard and compatibility modes. It resolves the actual mapped kernel32 x86 JMP, enters the production dispatcher through normal x86 execution, and verifies guest return/stack cleanup. Independent native x86 architecture and poisoned PBOOL assertions distinguish success from FALSE output, enforce the exact four-byte write with surrounding canaries, and cover repeated queries, an unaligned x86 destination, pseudo and actual OpenProcess handles, fabricated identities, valid child/alias queries, XP access denial, closed aliases, unwritable destinations and unknown-export rejection. Host process tests assert the exact query operation/right and retain authority failure controls. No setup was used as a regression retry.

Both regression modules and the failing pre-repair module are retained with source manifests. repair-fullbuild.log passes the required locked full build, including API/hash/dispatch freshness, handler stack cleanup, kernel32 image freshness and compiled data overlaps. Both final shipped repair modules are retained. Exact Git SHA, relevant binary diff and patches for owned new test and finding files preserve source recovery without archiving the checkout.

Known static imports GetPrivateProfileSectionNamesA and SystemTimeToTzSpecificLocalTime remain unsupported until actual reach proves their effect. Prior SFC ordinal 8/9 and SRSetRestorePoint lookup failures continued; no protection behavior is changed or fabricated. Runner exit zero was not treated as installation success. Whether the new repair permits installation remains unknown. The next useful proof requires a separately authorized genuine setup progression using this actual final target, followed only on real success by requested restart/provider checks and ordinary original AoE3 gameplay. This task used its one session and did not retry.

Final independent cleanup.json passes: 14 recorded owned PIDs absent, control port 19765 refused, no profiles created, four locks free, zero heavy processes, watcher absent and more than 5 GB free on /home. The check before publication is also retained as cleanup-before-publication.json. Box ID, state and expiry are N/A; coordinator lifecycle ownership is unchanged. No platform operation was rejected. The evidence containment checker exits zero; AoE3 has no corpus candidate entry, so these findings remain in the investigation note. The exact FINAL board entry releases runtime and all owned files after the repair and findings are published to main.