Donuts targeting and confirmation boundary, 18 November 2026

The native-scene controller now admits delivery while the reviewed target is fresh, without reserving another 500 ms inside targeting freshness. The same immutable targeting expiry still fences server preparation, physical application, native dequeue/validation and owner USER-return consumption. A receipt at or after that expiry is rejected permanently. Confirmation remains limited to 500 ms from delivery, physical UP to 1750 ms, and full operation/observation to three seconds. Legacy routes retain the conservative confirmation reserve.

Evidence: scratch/runs/20261118-donuts-targeting-consumption-confirmation-boundary/.

The applicable contract is the owner's explicit separation of native consumption from confirmation. The LMarbles distinct-deadlines and recovered-headroom notes were read alongside its older admission notes. Historical helpers reserved confirmation inside freshness, but supplied no independent authority requiring that extra reserve. LMarbles admission notes provide the completion allowance and independent-release design reference; they do not authorize a Donuts change. Applicability here rests on the actual Donuts native-scene transport and checks in lib/cli-guarded-input.js: valid, consume, dequeued, nativeValid, nativeConsumed and delivered. Queued stale effects are purged; release uses the original owner independently of freshness. Native receipts now additionally require the exact owner, epoch, nonce, deadline and owner USER-return boundary. The client keeps its strict consumption timestamp check and sticky terminal latch. As the LMarbles notes explain, a pre-effect check cannot undo an effect if a callback straddles expiry. Such a late receipt is never accepted or replayed; the independent UP is still awaited. This change supplies no exception to that receipt fence.

The old helper refused an actual elapsed-age crossing control before DOWN. The corrected helper consumed the compiled authored fixture before the identical five-second expiry while confirmation crossed that expiry within its separate 500 ms delivery bound. No capture or review timestamp was moved. Negative controls cover arrival after expiry, native queuing until expiry, actual native focus change, confirmation overrun, independently delayed UP, operation overrun, lost ACK with independent UP, and zero later controller endpoint calls after closure. The queued-expiry fixture remains alive past its real ten-second pump; its exact DOWN is purged and no native-return candidate is recorded.

An immutable published-source native fixture also passed the personal workflow. The whole 1024x712 desktop and its PUBLIC MENU were personally reviewed before one finite Right. Consumption occurred 3.973 seconds after capture, confirmation at 3.983 seconds, and independent UP at 4.094 seconds. Some earlier fixture captures displayed only teal despite live window metadata; these are retained as limitations, not treated as reviewed visible menus. Explicit authored-fixture repaint requests preceded a subsequent complete visible frame. They do not alter the original or supply original gameplay evidence.

The single authorized original reached LEVEL001 and accepted five finite gameplay gestures: Space, Right, Up, Space and Right. The post-turn frame visibly shows the ship facing right. Later frames show different positions, but motion was not isolated from respawn. These are actual consumed controls, but their score changes do not isolate a shot's consequence: score and fragments also changed while all preceding gameplay intentions were NOT_SENT. A last fragment remained in reviewed frames. LEVEL001 clear and LEVEL002 are unproved. Gameplay stopped at the fixed Quit reserve.

Ordinary Escape then consumed through the owner's USER-return boundary and the native exited0 with both EOFs, without host cancellation or watchdog signals. The existing final-choice path independently verified that exact owner's actual normal wait after release found the closed port. Physical UP was unacknowledged; it is recorded as OWNER_EXITED, never fabricated as an UP ACK. No later endpoint request, capture, rebind or replay followed terminal closure.

Shorter manual invocation permitted five gameplay gestures, although other complete-frame reviews still expired before DOWN. Whole-frame display at reduced scale was also tried, with native PNG hashes/timestamps unchanged. A duplicate review-file refusal before any input was followed by a new capture and intention; it was not an ambiguous transport result. Earlier black title captures were honestly reviewed as black; one finite Enter used the title's already-observed documented Begin Play route and produced the visible LEVEL001 frame.

The unchanged EXE, 37-file manifest, presentCap60 and normal guest Worker used sealed published source95997a6a5. Canonical module eb2395e1c115980dd78db646a26f755a4119a2a1e086942e715240e07c34bc6 was compiled from published e89057abd; the complete src closure is identical at the original source commit. Actual loaded module/source hashes are verified for the main thread and Workers. Preparation and native renderer/PNG preflight ran before the clock. Prior disk-stopped and UNKNOWN scopes stay closed. DirectDraw/Canvas2D does not prove a GL backend.

All eight final focused native cases and the locked build passed. One earlier near-expiry diagnostic safely refused while compiler load spent its prearm margin; the test expected a postdispatch fault and failed. A run-local manifest writer also encountered the append-only save rule. These failures are retained; the corrected manifest writer and final sequential suite pass. Automated tests pin the compiled module but use the live checkout, whose relevant patches are retained; the personal fixture and original instead use verified sealed sources. Both sealed launchers' invoking tool waits returned143 with unknown sender, despite their own completed normal owner/native/watchdog waits and descriptor closure receipts. Separate invoking-tool EOFs are not established. This limitation is not replaced with the inner native exit0 or a fabricated normal outer wait.

Native-supervise supplies finite process containment; independent watchdogs monitor actual home disk above 1.5 GB. No RAM gate, Python, Wine, box, worker, deployment, source archive or media copy is used. Git plus relevant patches is the recovery source. Selected whole PNGs, unique inputs/modules, actual waits and media hashes remain in the run; result.json is written last.