Native input consumer confirmation, 2026-11-17

Evidence: scratch/runs/20261117-caesar3-native-consumer-confirmation-repair/result.json. Git is the source archive; the run retains relevant patches and exact module identities.

The published baseline incorrectly acknowledged a guarded edge inside the host check_input import, before native USER filtering and owner routing completed. The same authored real x86 program, with an actual CPUWorker, proves false confirmations from an excluding PeekMessage filter and a sibling thread. Neither returned the edge to its native owner. Host application and queue removal alone are insufficient evidence of native consumption.

The repair issues an opaque token for the exact queued object and immutable intent. Native GetMessage/PeekMessage carry it through pending input and shared queue flags. Before effects, validation checks the original owner, memory, window/device identities, context and deadline. Confirmation requires the actual owner's successful native USER return, completed stack frame, no fresh yield, and matching HWND/message/wParam/lParam. The receipt includes the same nonce, token, owner thread, API and removal provenance. A successful owner PM_NOREMOVE return is valid raw-message consumption; it does not prove removal or dispatch. Filtered/sibling polling, empty polling, asynchronous key state alone and unrelated messages cannot confirm an intent. Browser/unarmed hosts issue no token.

Broader controls exposed a second concrete defect: ordinary resume_message_wait cleared reason 7 but could retain the old yield flag before re-entering GetMessage. The native handler could return EAX=1 with its complete stdcall frame while inheriting that old flag, losing confirmation. Resumption now clears only the old wait flag before the existing handler call. New callback or handler yields remain refused. No additional guest call, service flush, interpreter budget or deadline was introduced.

Confirmation lifetime and native MSG lifetime are distinct. Once a successful Peek returned an edge, its cached native MSG remains retrievable by ordinary Get across the confirmation deadline and independent physical UP. This cannot issue a second confirmation, change the original timestamp, or survive memory/owner retirement. Unconfirmed expired or cancelled edges are discarded before a guest MSG write. Invalid PM_NOREMOVE cleanup matches the exact token under the queue lock, so concurrent window purging cannot make it remove an unrelated successor. The authored delayed-Peek/Get control tests this with native execution.

The authored controls exercise ordinary title-style Peek/Get and pointer paths, Get-only, PM_NOREMOVE, delayed removal, excluding filters, actual asynchronous GetAsyncKeyState (0x8001), a nonconsumer, sibling polling, and bounded busy interpreter work. Independent physical release and watchdogs are prearmed; actual process waits and both EOFs are retained. Diagnostic host cleanup is explicitly not native Quit. Existing queue/device/window/generation/focus/modal, nonce/fresh-image, cancellation, retirement and transport refusals remain required. The metadata-scope test now states handler application without claiming native consumption while it deliberately suspends native polling.

The first original run's no-dequeue timeout remains unexplained. Its host handler applied DOWN at 59 ms, no native message was observed, cancellation occurred at 500 ms, and independent physical UP occurred at 1419 ms. A bounded authored GetTickCount workload reproduces that timing shape while its real CPUWorker progresses; this is a discriminator, not evidence that the original used that cause. Moving confirmation later cannot repair a missing native dequeue. Therefore no fresh original was launched, the old UNKNOWN endpoint was untouched, and title dismissal/city/Victory/Assignment2 food/native Quit remain unproved. Further original work requires a material causal timing repair and full exact CURRENT published Git qualification under the unchanged manual contract.

The media helper now inventories all files recursively before any future native clock and compares the complete paths, sizes and SHA-256 inventory afterward. The current 114-file inventory matches the retained postrun reference, but that reference is not retroactive pre-run proof. No media, EXE, defaults, save or clock was edited. No box, worker or public deployment was created.

Working sealed qualification passed all seven native pointer mappings, keyboard, ten consumer cases, ten identity/lifetime unit cases, the exact reviewed controller with authored native Quit, and permanent UNKNOWN with zero later requests. The real broker scope test also checks token isolation on main and worker routes. Canonical build, focused checks and final published Git/module identities are recorded in the result. Authored Quit is not an original-game Quit claim.