Sims type-library and child-export boundary, 2026-11-15

Evidence: scratch/runs/20261115-sims-typelib-child-export-boundary/. Predecessor order is TIMcritical994115, Q3schedule1037509, then this worker. Both predecessors released, and personal actual-wait/both-EOF/PID-group checks passed before existing source changes or any native run. No foreign WIP is adopted.

The previous ordinary original is permanently closed. Feature df42403751a84e476f6f66d957b4930c28c145df and report 3d087339589c66c790f433c4d256112f8e053cad establish one actual TreatAs mapping returning S_OK, child exit 0x80029c4a, rejected child export, no completion record or merge, and parent 0x80040154. Consumed titlebar input and physical UP did not establish dismissal or native Quit. No earlier attempt is retried.

Read-only findings

The retained child log explicitly loads genuine OLEAUT32 at guest 0x738000 and patches IKernel imports to it. Thus the unconditional WAT LoadTypeLib failure is not evidence of the actual failing operation. The old filtered API trace omits the type-library/resource/file operations needed to discriminate.

Pinned IKernel SHA-256 is 985eb55ecb750da812876b8569d5f1999a30a24bcc54f9bab4d3fc44dfedb931. Its OLEAUT32 IAT entries are 0x46c284 RegisterTypeLib (ordinal 163), 0x46c288 LoadTypeLib (161), and 0x46c2b0 LoadRegTypeLib (162). Its PE has named TYPELIB resource ID 1, language 0x409, RVA 0x9d1f8, size 0xd8ae, format SLTG, resource SHA-256 475c366495b0a699277ab0fcfab047a9ce9e043b1d8c6b30c3d2ccdb5841baae. Resource presence alone does not prove selection or loading.

Static helper 0x432423 obtains the module path through GetModuleFileNameA, converts it to wide text, and calls LoadTypeLib at 0x4324db. On failure it replaces the extension with .tlb, converts again, and calls at 0x43256a. Its caller 0x432591 calls RegisterTypeLib at 0x43260b only after success, then releases the returned interface. Other static LoadTypeLib callsites exist; this describes possible control flow, not observed original execution.

Microsoft LoadTypeLib specifies standalone files and PE resources, default first resource, optional numeric resource suffix, reference counting on repeated loads, and implicit registration for names without a path. LoadRegTypeLib selects registered versions and delegates loading to LoadTypeLib. RegisterTypeLib persists library/interface registration, not coclass CLSID registration. No fabricated ITypeLib pointer, HRESULT success, or installation state is valid.

The child completion exporter currently skips provider-backed entries and rejects every resident key outside canonical C paths. Its error omits the key. The general parent snapshot exporter strips only a C prefix, so a resident non-C entry can be represented under a colon-bearing host directory. Whether this produced the actual rejected key remains unknown until diagnostic evidence. Any diagnostic must preserve rejection and leave completion absent.

The new pure-host candidate passes 17 adversarial keys, lazy-provider exclusion, readonly-drive rejection, and nested failure without partial output. A synthetic resident D file exported with the real general exporter and remounted through the real include-path producer/VirtualFS retains its D key and is rejected by the completion boundary. This proves that producer mechanism is possible; the actual rejected original key is still unknown. No producer change is justified from this control alone.

The SLTG data contains the literal imported-library reference *\G{00020430-0000-0000-C000-000000000046}#1.0#0#C:\WINNT\System32\StdOle32.tlb#. Current system-data/default registry entries provide stdole2 version 2.0. The Microsoft LoadRegTypeLib contract requires the requested major version; version 2.0 alone does not establish availability of 1.0. Whether this import caused the old original remains an inference pending a fresh discriminating trace. The generated controls separately test version lookup, genuine resource loading, registration, and registration after the genuine standard dependency is supplied through real APIs.

The existing Windows 98 SE ISO scratch/media/win98se-oem-x05-29232.iso (SHA-256 2adfb46df8a9c7bbd2f67bff07461cc2f9d9ec8e01f0e112cb044c9e3e62f607) contains win98/WIN98_52.CAB:stdole32.tlb, 7,168 bytes, SHA-256 308e70310b4707f63095db73e67ba080cb67657abdcba7efe66f0f7f5ce017ed. The new native test includes a pinned recovery recipe. Only this data member was retained as a generated fixture input; no cabinet/media copy or guest execution was retained. Runtime defaults now mount this exact local-only data and seed only the standard library 1.0 path beside the existing 2.0 path. Recovery: node test/test-sims-typelib-native.cjs --recover-standard test/binaries/tlbs/stdole32.tlb uses the pinned existing ISO/cabinet/member; the ignored binary is not committed.

TIM's explicit release was personally checked against its sealed closure: 123 actual waits with both EOFs, including direct native/watchdog/supervisor/ launcher roles, and 124 currently absent PIDs and matching groups. Q3 is next; all preparation here remains LIGHT until Q3 also explicitly releases.

Planned discriminating qualification

Generate ordinary native clients with the pinned genuine TYPELIB bytes as resource/file data only. Never execute the unpacked IKernel or game as a substitute original. Trace genuine OLEAUT32's file/resource/API boundaries and HRESULT consequences. Qualify exact NULL-owner MB_ICONHAND|MB_OK client OK, separate Cancel controls, and identity/dismissal/exit consequences. Reproduce nonclient close before considering any hit-testing change.

Add bounded escaped rejected-key and entry/drive provenance diagnostics, with absolute/traversal/drive/UNC/NUL/device/readonly/nested adversarial controls. Repair a producer only when demonstrated. Keep path/completion/merge and all input ownership guards strict. One new unchanged ordinary original is allowed only after a material repair or necessary diagnostics qualify and publish.

Native causal controls and repair

The generated client embeds only the pinned IKernel TYPELIB data and calls actual genuine OLEAUT32. Its baseline LoadTypeLib succeeds, yields 87 type infos, and repeated loading/release respects reference counts. RegisterTypeLib then returns 0x80029c4a: the actual registry trace requests standard GUID {00020430-0000-0000-C000-000000000046} version 1.0, finds only 2.0, and fails its StdOle32.tlb fallback search. Explicitly registering the pinned standard data first makes IKernel-data registration and registered reload succeed. The ordinary default data/path repair makes all eight generated controls pass: self resource, standalone, indexed resource, missing-file failure/NULL, registered standard 1.0/2.0, direct standard 1.0, and explicit standard-first registration. Final interface release reaches zero. This proves the dependency failure in the native control, not yet the old original's exact operation. The one newly authorized original will trace the genuine IKernel registration helper and actual registry/file selection.

Exact NULL-owner MB_ICONHAND|MB_OK client OK returns IDOK and exits natively. A physical titlebar X baseline is consumed with timely physical UP but leaves the same HWND/generation. Read-only fixture inspection establishes HTCLOSE, the ordinary native nonclient release route, and the resulting queued WM_NCLBUTTONDOWN/HTCLOSE. Adding MessageBox procedure close translation alone still leaves that message queued and the same dialog alive. The parked modal pump lacked posted-message dispatch. The repair dispatches one posted message filtered to the owning built-in MessageBox HWND, without entering guest dialog procedures or adding a continuation target. Its procedure translates HTCLOSE through SC_CLOSE/WM_CLOSE and chooses the actual Cancel or sole OK button result. No hit-testing, renderer/input/controller, source guard, focus/capture, thread/process/HWND-generation rule changes. Final qualification passes 19 exact-error native tests and 18 NULL-owner OKCANCEL native tests, including physical OK, X and Cancel with actual guest exit zero, no host cancellation, and physical UP in 202–269 ms. Both shared and committed-HEAD plus whole-owned full flock builds pass. Adjacent native modal/shadow regressions and publication checks precede the original. A deliberately wrapped fixture input method was correctly refused by the source guard and permanently closed; that fixture was never replayed.

Worker-shadow MessageBox, reentrant common-modal continuation and MessageBox A/W regressions pass. The optional older IDNO fixture cannot compile because it names removed global $esp; the same failure is reproduced on untouched committed HEAD, before the repair. This is a stale fixture limitation, not a passing regression or a full-build failure. Both full builds remain green.

The unsafe completion guard is unchanged. Its bounded ASCII-escaped diagnostic records up to 256 UTF-16 key units, total key length, resident provenance and readonly-drive status. Adversarial absolute/traversal/drive/UNC/NUL/device, readonly and nested controls still reject without completion, registry output, partial write or merge. No producer normalization is changed without an actual rejected original key. A fresh original is justified by the qualified standard library repair and exact native dismissal repair, with additional discriminating helper/API traces; it is not an unchanged retry.

One ordinary original, permanently closed

Feature b95d0c8bc6fd128af1c1b56befe15f04c18b5ee2 was committed on main and strict-pulled/pushed before the one fresh authorized ownerNova original. Exact Worker-safe preloads and real authority Workers pass; removing either source or host-map hook is refused. The qualified module/source match was verified against that published commit. An independent watchdog was armed before launch: 900-second native limit, 920-second independent deadline, 180-second Quit reserve, real ticks and unchanged original defaults.

The original's genuine child trace now shows LoadTypeLib returning S_OK at 0x4324e1, without taking the .tlb fallback, then RegisterTypeLib returning S_OK at 0x432611. Genuine OLEAUT32 reads the installed IKernel PE resource and C:\Windows\System\stdole32.tlb. Its ordinary -RegServer child completes one real TreatAs call and exits zero, advancing past the previous typelib failure. The newly bounded diagnostic rejects exactly d:\maxis.ini, length 12, resident provenance, readonly-drive false. No completion record or child state is merged; parent still shows 0x80040154. This is not installed state or a successful ordinary installation.

The stored full images show the fatal error, native main and same-process, same-thread NULL-owner native MessageBox. During the run the conversation review of later images appeared teal, so no current visible target authorized input. That review failure was written UNKNOWN, with zero binding/input gestures and zero later control requests, and permanently closed by an orderly supervisor host stop. Subsequent offline review of the unchanged PNG base64 with explicit image/png MIME correctly shows the error dialog. All four original PNGs have the same SHA-256; blank native pixels are not proved. The earlier view_image data URL used application/octet-stream. This correction does not reopen the run or authorize input. Actual harness wait zero is not native Quit. No original is retried. Native OK/X/Cancel fixture exits remain separate from this original outcome. A newly authorized run needs correct image MIME and fresh current-scene review before any genuine installer UI gesture.

Actual parent/child consumed-source pins match the published commit; every nonempty parent/child/Worker mapped-host receipt is hash-verified. All 3,172 launch pins match after closure, including original Setup, exact nativepartial overlay index and 614,532-byte IKernel. Personal direct closure establishes 399 actual waits with both EOFs, 622 recorded PIDs/groups absent, and both locks free before any follow-up source edits. No box was created.

Readonly profile producer

The actual offending file is the existing disc's 16-byte root maxis.ini, SHA-256 d00adbd56f1dc99a9bf5652035f0b56e8531e31aea7d9ebe25e512011aaed7ed. The real media mount plan eagerly materializes INI/LID files for synchronous profile APIs. Baseline materializeEntry assigns through the provider data setter, which discards provider provenance even on a read-only drive. A direct read has the same promotion. The general parent snapshot therefore emits d:/maxis.ini; the real include-tree producer remounts it as resident D data without the source drive's readonly metadata, and completion correctly rejects it. The pinned-media test reproduces this exact producer before the repair.

The filesystem repair caches reads from read-only drives without relinquishing the borrowed provider. Its private nonenumerable cache setter is used only by successful async materialization after the existing source-identity/range check. The ordinary data setter remains a write. Reads and copies on writable drives keep their prior resident behavior, even for files with the readonly attribute; changing a drive back to writable likewise restores that behavior. Adoption retargets each provider read policy to the new active filesystem; a focused ownership-transfer test caught and prevents consulting the old drive policy. The child completion guard is unchanged, including direct rejection of untrusted resident D entries on a readonly drive. No path is sanitized, made trusted by its name, or manufactured into a completion record.

A generated native parent on a generated 60-KB ISO reads the readonly INI, uses ordinary ReadFile/WriteFile to stage its generated child on C, spawns it, waits, and checks the child's actual C-file and registry value. The old producer reproduces the exact d:\maxis.ini rejection and parent exit 25. The repaired control exits zero in parent and child, with one real file and 78 registry/INI keys merged; the parent verifies its child-written file bytes and DWORD. VFS, lazy-entry, export, adoption and media-autorun regressions pass, as do shared and committed-HEAD-plus-whole-owned full flock builds. The media CLI suite contains a stale source-text assertion for setExeDrive; it fails identically without the filesystem repair and is retained as a baseline limitation. This is a native control, not another original or an installed-game claim. Two earlier fixture setup errors (omitted explicit EXE, then a provider-backed unstaged child EXE) are retained as unsuccessful preparation diagnostics and do not count as causal controls.

The inherited-D-cwd generated control also exposes a separate existing boundary: when all source files remain borrowed, file-only child snapshots do not carry empty source directories. Its child refuses the absent d:\ working directory; the independent watchdog drained this unsuccessful control. This is retained as a next-boundary limitation, not a successful merge. The positive producer control passes explicit existing C:\ as the ordinary CreateProcess working directory, while still reading the readonly D profile and preserving the same completion guard. No source directory or installed state is fabricated. Propagating genuine readonly source directories/media to children remains separate necessary work before claiming ordinary installation can proceed beyond this boundary.

The explicit-C-cwd control isolates mutable file/registry transport only. It does not qualify the genuine inherited-readonly-D-cwd Sims route. The concrete handoff is to preserve the genuine readonly mount/directory context across CreateProcess without exporting cached disc bytes or fabricating D state, then prove that exact inherited-D-cwd parent/child control. The current task leaves that boundary unresolved and consumes no further original authorization.