Neverhood corrected-registry installer admission: native BUTTON owner boundary

Evidence: scratch/runs/20261117-neverhood-corrected-registry-native-install-gameplay/result.json.

No fresh original installer or installed shortcut was launched. Published registry repair 42feb3f564442df3a7344f7b20162def51216a3a remains inherited. The exact published whole source tested here is e7a7e7b49ae28dd94b5bcffc0245cde0a0324351; no shared uncommitted source was adopted. The source inventory, compiled production module and helper preload were sealed before tests. Actual Node main, Worker and OS-child inheritance matched.

The five retained ordinary installer actions all have real native BUTTON child metadata and native-button-dispatch-return receipts. This is narrower than an arbitrary mouse/key route. The installer could potentially avoid the queued USER route, but those old receipts do not qualify its current published runtime. Read-only provenance again verifies full 257-member CD, genuine DLLs, all 58 retained overlay blobs and the genuine installed shortcut target setup95.exe. The old final registry still lacks the three installation paths; it was never modified, reconstructed, hydrated or launched.

Current published storage, controller, provenance/tamper, typed BUTTON unit and seven native pointer/presentation cases pass. A diagnostic build of that same Git closure also passes all 92 native registry cases, including actual CPUWorker and OS-child persistence/merge controls. The latter include real scaled BUTTON routing, clipped-child refusal and changed-viewport refusal, independent release and actual native/watchdog waits. They do not supply a positive CPUWorker-owned BUTTON or normal installer-exit identity qualification.

A required wrong-owner negative control fails on the actual production module. An authored x86 executable creates the actual native DialogBox and BUTTON, owned by thread 1. The published captureNativeButton captures that real actor and its sequence 0. Within one authored diagnostic call, the native set_current_thread_id(99) fault changes the executing actor before the real dialog_route_mouse_screen call. Routing returns 1 and the native sequence advances to 1. Published confirmNativeButton accepts the receipt as thread 1 although the executing actor is 99. No JavaScript receipt counter or target was fabricated. This is an actual native owner fault injection after capture, not an assertion that ordinary original execution changes its actor this way. The detailed record is wrong-owner-main-qualification/button-callback/wrong-owner-proof.json.

The cause is visible in published source: dialog_route_mouse records the thread from wnd_get_thread(ch), rather than validating that it is the executing actor; confirmation compares that recorded owner with the captured target. Neither the receipt fields nor that confirmation pin the native invocation identity across this dispatch. The pre-capture actor check alone does not cover the post-capture negative control. Returning from the WAT control dispatch also must not be presented as completion of a later queued guest callback.

Two earlier authored nested-focus/owner-draw probes generated later nested callbacks and bounded-send abandonment logs. Their CONSUMED receipts precede those callback logs. They establish no false callback-return acknowledgement and are not counted as a causal failure. An initial probe targeted an outer button after its setup had already opened a nested dialog; its bind refusal is retained as an invalid experiment. A subsequent probe referred to an unavailable eval variable and is likewise not a qualification result.

The first concrete boundary is therefore failure of the required native BUTTON wrong-owner control, before any original admission. The relevant WAT, native USER helper and export files remain under VB/C3 whole-file leases; this worker did not edit them or weaken the guard. Their owners need an execution-owner and invocation-fenced native receipt/refusal, with exact-current main/CPUWorker, modal-generation, nested-callback, fault/cancel and normal-exit identity tests. The queued USER repair remains independently required for game input. No original retry, old-endpoint query, registry fabrication, restart or deployment was performed. No corrected-install, Play, room, object action or native Quit is claimed. All authored native processes were joined after explicit host cleanup; that cleanup is not native Quit. No box was created.

Recovery uses the exact Git SHA, retained relevant binary patch, manifest hashes, small authored probe inputs and the production module. Media is referenced by its existing path and SHA-256; no source or media archive was made.