AoE3 repaired DirectX setup — 2026-11-05

The one genuine setup session progressed past the repaired administrator query and entered genuine DSETUP, then failed because kernel32!GetNativeSystemInfo was absent. That export is now implemented using the emulator's existing native x86 system model and covered by a regression. Setup has not been retried with this second repair. Completed DirectX installation and original player-controlled AoE3 gameplay remain unproved.

The predecessor repaired the earlier missing export, advpack!IsNTAdmin, in fb256f9f6456e1d790e9fe0813adaec28cdadc35. Its genuine setup failed before DSETUP entry, wrote two logs, and left 73 registry entries unchanged. There is no setup-produced installed DirectX state to borrow or claim.

Evidence is scratch/runs/20261105-aoe3-repaired-dxsetup/ (ignored). The single XP-profile DXSETUP.exe /silent session reused the existing complete 146-file setup at scratch/game-media/aoe3-dx9c-20261105/feb2010-complete/ and the predecessor's actual final registry and three-record overlay. Only its two tiny actual logs and directory metadata were seeded into the owned overlay; the base media was mounted in place. No extraction, download or repeated whole-media verification occurred. The redist pin remains f6d191e89a963d7cca34f169d30f49eab99c1ed3bb92da73ec43617caaa1e93f; current trust/revocation remain unknown.

Gate, closure and actual module

Preparation remained source-only until Conan1935 FINAL 18:16:14Z, Aranna1945 FINAL 18:31:58Z and legacyADD1955 FINAL 18:41:45Z. Independent release-gate.json passed: 353 recorded PIDs absent, 145 ports refused, 38 profiles absent, four locks free, no heavy processes and free disk above 5 GB.

Cheap current boot/font/native-DLL closure checks and the exact import audit preceded a successful full build under flock /tmp/wine-build.lock. The three setup-module hashes match the existing pins. COMCTL32 ordinal 17 retains its InitCommonControls mapping. The stricter import audit found two unsupported dsetup32 static imports omitted by the predecessor's summary: GetPrivateProfileSectionNamesA and SystemTimeToTzSpecificLocalTime. These remain explicit gaps; neither was the observed terminal failure. The initial overly strict audit assertion and the corrected audit are both retained.

isntadmin-repair-source.json proves the handler exactly matches fb256f9f…; the fresh include closure also contains its dispatcher call, and both retained shipped binaries contain the correct API hash/ID row. Actual setup used main d356b2505f99499673b4652bde6adb7afdf642fc, with an empty relevant tracked production diff. The standard module was 2173611 bytes, SHA-256 144d156bec9e4be2654fafeb67d9bf9b6746d514ee83348e2f401547aa7ff4a4. tested-modules.json, runtime-source-hashes.json, the relevant binary diff and small untracked-source patch identify the inputs. Git is the source archive; no source-tree archive or repeated media copy was made.

Genuine progression and first failures

setup-process.json records the exact production-runner command, PID 3240945, and host execution 2026-11-05T18:43:55.531196Z–18:43:56.300699Z. The run-owned adapter records frozen baseline/final inventories and suppresses bulk VFS export copies; the production writable overlay journals actual writes. The predecessor's 73-entry final registry was imported once before execution, never restored afterward. /silent is the previously documented option; silent license handling was not visually observed.

Both actual administrator queries return true. The passive counter records one hit at genuine DSETUP!DirectXSetupA 0x0118806d, compared with zero in the predecessor. Actual logs show DSETUP's entry and SetupForDirectX() starting, then the installer's own Installing on WinXP SP2 branch.

The first new lookup failures are SFC ordinal 8, ordinal 9 and SRSetRestorePoint (API calls 2013–2015). Setup logs LoadSfcDLL() failure and continues. Those are preserved as the first observed misses; no SFC or restore/protection behavior was changed. The subsequent terminal failure is API 2375: GetProcAddress(kernel32, "GetNativeSystemInfo") returns zero. The actual error log reports error 127, then IsIA64(): Unable to initialize CKernel32. Setup finalizes phase 0 with value -9, internal/unsupported error. The guest calls ExitProcess(0xfffffff7) (-9); host runner exit zero is not installation success.

Only the inherited logs changed: directx.log is now 1356 bytes, SHA-256 ef568d1aaa460494103d47cb79b98d98c31ef17969feb3a61399ab8ca0291f14; dxerror.log is 1116 bytes, SHA-256 98f446ca7bcaa84e7c9616b3671416142c32c73095de56c1f7f598dfa0da3a5d. Their initial 308/464 bytes are the predecessor's unchanged failed-admin messages; the added suffixes are preserved separately. Guest dates use the pinned June 1999 clock, while the host timestamps above identify this session. Overlay metadata and actual bytes remain in target-overlay/.

setup-file-delta.json records no other changed or deleted file. Registry baseline/final both have 73 identical entries, with no pending replacement, registration or restart delta. No restart or provider check was performed, and no original AoE3 session ran because setup failed before installation. I personally inspected setup-final.png: a uniform teal 640×480 desktop, without a completion dialog or gameplay.

Repair and validation

Microsoft documents GetNativeSystemInfo as a void kernel32 function, supported from XP, that fills SYSTEM_INFO. This emulator's installed OS model is native 32-bit x86 without WOW64. Its new handler delegates to the existing GetSystemInfo behavior, including the 36-byte structure and one-argument stdcall cleanup. It does not report the Linux/JavaScript host's architecture. API ID 4603 is appended; the explicit kernel32 PE export is added at next-free emulator ABI ordinal 596, preserving all existing ordinals. The kernel32 image, API hash and dispatch outputs are regenerated through the existing tools. No version, DirectX capability, AttemptToRun, installation or security state is synthesized.

test/test-get-native-system-info.js first reproduced the absent dynamic export. A handler/table-only intermediate still failed because kernel32's explicit PE export surface also needed the new entry; that failure is retained. The final regression passes in standard and compatibility modes. It uses the real PE loader, dynamic kernel32 lookup, mapped export JMP, production dispatcher and actual x86 entry/return path. Independent native x86 structure assertions, poisoned destinations, boundary canaries, repeated queries, unchanged GetSystemInfo, unknown-export rejection and exact stack cleanup prevent a resolve-only or constant-success implementation from passing. The exact tested harness modules and source manifests are retained. One artifact-preparation command initially used a relative path and was corrected to the helper's required absolute path; no guest ran from that preparation.

The optional older test-kernel32-image.js excerpt fixture fails on omitted loadlib_builtin_only; a run-owned selection-only adaptation exposes a second omitted loadlib_name_is_qualified. Both failures are retained. Those helpers are unchanged production code, and the shared fixture was not edited. The new regression instead covers the full production loader and actual export entry. The final flock /tmp/wine-build.lock bash tools/build.sh passed at 18:51Z, including the API/hash/dispatch, handler-stack, kernel32-image freshness and compiled data-overlap gates. repair-fullbuild.log and repair-tested-modules.json retain the result and both shipped repair binaries.

The next required proof is a separately authorized genuine setup progression with this new export, retaining the same actual target state. SFC/restore misses and the two static import gaps remain unresolved surfaces; whether any becomes fatal later is unknown. This task used its one setup session and did not retry. Genuine completion, any requested restart, resulting provider state and then original player-controlled AoE3 gameplay remain conditional.

Final independent cleanup.json passed: 21 recorded owned PIDs absent, control port 19765 refused, no profiles created, four locks free, zero heavy processes, and 7622541312 bytes free on /home. The board watcher was terminated and is absent. Box ID/state/expiry are N/A. No box, subworker, public deployment, protection change or laptop Heroes II operation occurred; coordinator lifecycle ownership is unchanged. No platform rejection occurred. The repair and this finding are published together through explicit owned paths; the corresponding exact FINAL board entry releases runtime and every owned file after publication.