GOTY authored dialog display preparation, 2026-11-18
Evidence: scratch/runs/20261118-ut-goty-native-dialog-display-install/.
No original installer or game has run at this preparation checkpoint.
The previously reported blank capture has SHA256
8b6859eb449d3228524b27650c5ce110fca37c65853867dbc838013d2c142642.
Fresh exact authored MessageBox/read-only-disc/self-ShellExecute controls at
1024x712 produce that same complete PNG, including 52,080 non-desktop pixels.
The original compressed bytes decode to a caption, message, focused OK and
Cancel. Repeated captures are byte-identical. Treating this hash as evidence
of a blank native render was incorrect.
Existing --trace-ctrl, --trace-gdi, --trace-composite and generic host
tracing establish the actual path: native control paints into canonical DIB
0x1c000000, surface 0x610001, attached to live HWND 0x10001, with process
1000 and the matching instance. The compositor selects this window and its
back surface. Passive pre/post reads retain canonical and derived surface
checksums, ownership, geometry and clip state. No window, focus, clock,
invalidation, input fence or painter was synthesized to obtain these facts.
The pinned module is SHA256
514478fe584a69e23792575ab94ea497e859b9034d978e5d2ed943d0e8f49c98.
The locked full build reproduces it byte for byte. Git commit, relevant dirty
binary patch and source-file hashes accompany the retained module; sources
were not archived. The authored fixture was reconstructed from the retained
small C/JavaScript recipe after its original EXE disappeared.
The file-image tool returned an application/octet-stream data URL. An explicit
PNG MIME type and lossless re-encoding each produced a personally visible whole
dialog, but later repeated previews were still inconsistent. MIME alone is
not a proved complete repair. The remaining preview-transport cause is
UNKNOWN; there is no demonstrated native USER defect to patch. The earlier
unpinned scope stays closed and no causal module regression is claimed.
tools/ut-goty-dialog-display.cjs prepares an existing whole PNG for review:
it verifies the signature, CRC/decode and optional expected SHA256, supplies an
explicit image/png content block, and reports decoded dimensions and pixel
census. It preserves every original encoded byte. It never calls a runtime
endpoint or claims personal review from a pixel count. Regression controls
cover format, unchanged encoded/decoded bytes, SHA mismatch, truncation and
corruption. This repairs missing evidence validation in preparation, not USER
painting or the external image viewer.
A real Chrome control mounts the same authored read-only disc through normal media import and runs the unchanged authored PE in the normal guest Worker. Debugger bytecode readback matches the pinned module above. The complete host screenshot personally shows the caption, message and OK/Cancel; its whole guest canvas also contains exactly 52,080 non-desktop pixels. Browser input and child completion are not claimed. The first browser attempt's local bytecode-collection error was corrected with actual flattened Worker sessions; both browser processes have actual waits and both pipe EOFs.
One native investigation submission was mistakenly batched with image inspection and described a prior positive frame. Its native consumption, independent UP ACK and parent/child Exit0 remain evidence, but it is explicitly excluded from reviewed-input acceptance. A later zero-input diagnostic scope was cancelled. The fresh positive native control selected a 1024x768 screen before launch, was personally reviewed in a separate tool turn, and submitted one fixed 300ms Enter with the published guard helper's final-choice path. The original 500ms confirmation, 1750ms independent UP, identity/generation fences and 900/920/180 custody policy remain unchanged. Actual wait receipts, rather than a post-exit endpoint, determine completion.
At the initial local-media check, the unique recorded ISO was absent, as were
other workers' retained binaries. This worker performed no download or disc
copy. The retained Manifest permits a mandatory source-size census, but its
.uz map sizes do not establish expanded installation space. Exact expanded
space remains unknown until the unique disc is available. No installation
write or incomplete game launch is justified by that lower bound.
Ordinary bootstrap and provider-backed child repair
The coordinator subsequently restored the one missing disc at its original path,
with an independent whole-file rehash of all 649,633,792 bytes to the original
e184984ca88f001c5ddd52035d76cd64e266e26c74975161b5ed72366c74704f.
This worker rehashed it again and performed no download or duplicate disc copy.
The normal media importer selected D:\Setup.exe from AUTORUN.INF and extracted
only that process image for the native loader. The first actual original scope
exited without displaying a wizard or installing any files: both ordinary
CreateProcess and ShellExecute attempts to D:\System\Setup.exe failed with
"could not be exported". Parent/supervisor/watchdog have actual waits and EOFs.
A generated provider-backed PE control reproduces that failure with the same native runner and pinned module: the old snapshot-only lookup fails before any dialog and the guest exits 47. Source-disc entries deliberately retain their provider even after reads, so the mutable-state exporter correctly omits them. The selected child image now gets a separate temporary loader file, outside the mutable snapshot, while its original media ownership and mount context remain. No other media entry is read or exported. Unfilled asynchronous providers still fail explicitly; partial bytes never launch.
The first repair control exposed a second failure: seedExeImage replaced the
restored read-only executable with a resident entry, and the existing completion
guard rejected that unsafe D: state. Seeding now verifies matching loader bytes
and preserves the exact source-owned object on read-only drives. Missing owners
or mismatched bytes fail before alias writes. The completion guard is unchanged.
A fresh native control starts the actual D: child, validates read-only media and
exact renderer-test arguments, writes PASS, exits 0, and completes a guarded
merge. Its whole dialog was personally reviewed. A later Enter closed the parent
before independent UP could be acknowledged, so that submission is UNKNOWN and
is not an input acceptance pass; its scope stays closed. The earlier exact
self-child control remains the successful reviewed-input control.
The browser control also launches a real provider-backed D: child through the normal shell. Its whole host screenshot personally shows the child-authored message after checking read-only drive, profile and exact arguments, alongside both running apps. Actual Worker bytecode matches the pinned module. Browser input or completed child merge is not claimed. An earlier browser diagnostic showed real child Exit0 but parent marker timeout; that separate browser merge behavior was not changed in this native loader repair.
Focused checks cover snapshot exclusion, selected-image bytes, unchanged source identity/provider, read-only write rejection, unrelated lazy media exclusion, resident-image reuse, missing image, asynchronous-source rejection and immutable seed mismatch/missing-owner rejection. The existing VFS seeding suite passes.
Native private DLL discovery parity
After publishing the loader repair as 49dd59dba (strict pull and push both 0),
the second fresh ordinary autorun reached System\Setup.exe but failed at batch
0 on the decorated UClass constructor imported from Core.dll. The actual Core
DLL is on the disc; CLI graph admission considered only the registry's names,
so it never asked its existing guest-path resolver for this private sibling.
The browser already admits genuine mounted private siblings. A narrow native
admission helper now applies that same rule to imports: actual files in the
executable's guest directory, excluding builtin-only system DLLs, static module
placeholders and path-shaped requests. No registry list, app registration or
explicit DLL seed was changed.
A generated disc child importing PRIVATE.DLL reproduces the exact old failure:
DiscValue becomes an unimplemented API and the child crashes at batch 0.
With native admission, the actual sibling PE loads, DllMain runs, DiscValue
returns its authored 0x432 value, the child exits 0, and its PASS file merges
through the unchanged completion guard. The whole parent dialog was personally
reviewed in a fresh scope with a preselected separate control port. An earlier
fixture overlap caused EADDRINUSE and the identity guard refused capture; that
scope was cancelled without binding or input. No original scope was affected.
Focused negatives cover builtin exclusion, sibling-directory scope, explicit
path refusal, placeholders and admission without touching provider bytes.
The browser private-DLL control also personally shows the child-authored dialog;
its host log proves PRIVATE.DLL discovery, real PE mapping and DllMain return.
Both actual Workers' bytecode matches the pinned module. No browser input was
sent. Focused checks and the subsequent locked full build passed.
For direct review of the same genuine System\Setup.exe selected by the autorun
bootstrap, ordinary importer selection exposed a separate same-basename bug:
asking for its complete path returned the earlier D:\Setup.exe by basename.
chooseCandidate now permits basename fallback only for bare names. Complete
paths remain exact, including absent-path rejection; ordinary bare-name default
selection is preserved. This was detected before any guest launch. The focused
media CLI test also had a stale source assertion predating configureExeOwned;
its executable-drive precedence check now names the current owned boot API.
Both that suite and the same-basename regression checks pass. No manual
registry, INI, FirstRun, renderer setting or game patch is involved.
Retention uses selected named whole PNGs only. Byte-identical investigation
captures and derived lossless/BMP previews were removed after recording their
hashes and recovery aliases; selected-image-retention.json maps them to the
retained original whole capture. No source tree snapshot, frame series, capture
bundle or second original disc was made.
Final original outcome and blocker
Exact-path selection was published as 0ceafffe1 after focused media/child tests
and another successful locked full build, followed immediately by strict pull
and push. A third fresh original scope selected the genuine
D:\System\Setup.exe with no parameters and D:\ as its working directory, exactly
the image and inherited directory observed from the ordinary autorun bootstrap.
This used normal read-only media import, no registered app, overrides or patches.
The source and pinned module are recorded in installer-source-files.json plus
the relevant dirty-source binary patch.
The real original Window.dll, Core.dll, MSVCRT.dll and COMCTL32.dll loaded and ran their initializers. The original created its top-level setup dialog plus two nested child dialogs and called ShowWindow for the first wizard page. The reviewed-presentation endpoint then remained pending for its complete fixed 3000ms window, reporting main.topLevel as its sole failure. No whole original PNG was produced, and no original control was personally reviewed or bound. The underlying main-window classification/selection cause remains UNKNOWN; the creation log alone does not establish whether USER state or renderer metadata is wrong. No guard was suppressed and no main HWND, focus or window was replaced. No endpoint request followed the terminal capture failure. The scope was closed with orderly TERM, actual native/supervisor/watchdog waits and both pipe EOFs. TERM cleanup is not an ordinary installer Cancel or game Quit.
This is the concrete blocker to normal installation: a reviewed original wizard is unavailable. The mandatory manifest census remains a 207,876,925-byte source lower bound including 81 compressed map entries (50,289,750 bytes). Exact expanded mandatory space was intentionally left unknown; no Install action or original installation data write occurred. The original overlay reported zero records flushed. Three original installer scopes ran in total; zero original game scopes, 3D world, movement, fire or ordinary Quit are claimed. Earlier UNKNOWN scopes were never reopened. No Windows reference behavior was inferred from Wine.
All owned native runtime scopes have retained actual waits and EOFs. No test box was created: ID null, state NOT_CREATED, expiry null, stop receipt null. The final evidence index and cleanup verification accompany result.json, written last. The unique disc remains at its original recovery path and is referenced by whole hash; selected process images are recoverable through ordinary import and are removed from temporary loader storage after scope closure.
The final audit covers 18 runtime scopes and 88 distinct recorded PIDs: all native, supervisor, watchdog, original-child and browser Chrome wait/EOF receipts passed; no recorded process remained live. The final whole ISO rehash still matched the original SHA and /home free space was 4,035,854,336 bytes. An initial standalone board-follow helper did not retain a PID/wait receipt; its actual wait is not claimed, and no matching live tail remained in the final process scan. This limitation does not replace or weaken the retained native runtime wait evidence.