Donuts LEVEL001 continuation and native Quit attempt, 18 November 2026

The single authorized original reached its title, accepted ordinary native Enter, displayed LEVEL 001, and continued in a live arena. First-level clear/LEVEL002 and ordinary native Quit remain unproved. The concrete blocker was the selected controller's immutable five-second gameplay review window: reviewed Right, Space and Escape attempts were all refused before DOWN/prearm at ages 10.966, 7.277 and 5.477 seconds. No movement or fire key was delivered in this run. No guard was relaxed, original relaunched, or endpoint used after host cancellation.

Evidence: scratch/runs/20261118-donuts-original-level1-clear-native-quit/. This is a controller/operator freshness limitation, not a reproduced emulator crash or evidence that Donuts cannot clear the level. The earlier accepted turn/thrust/fire and score100 evidence, and the version-3 DirectInput Device2 repair, remain as described in graphics-donuts-current-crash-20261107.md. No runtime or API fix was made.

Personally reviewed whole frames

Five original 1024×712 PNGs were reviewed, with hashes and observations in visual-review.json:

The red top HUD reads 99 in arena frames; its meaning is not certified here. Neither fragments nor increasing score establish a level clear. After reviewing the last frame, ordinary Escape was attempted using the title's documented exit route, but refused at 5.477 seconds before any DOWN. The controller had already entered its irreversible Quit phase. Host cancellation then closed the scope. Native process exit0 under SIGTERM handling is cleanup, not native Quit. No fixed macro, guest/save/score/level edit, cheat, collision attribution, projectile trajectory or FPS claim is made.

Prepared identity and authored native controls

Runtime source is published main 585235294ac4fa98220dbb8cb39a94922e9bd95c. Sources were read as exact Git blobs into transient sealed descriptors before the original clock; dirty shared-tree changes were excluded. relevant-worktree.patch records those excluded relevant differences for audit. module-build.json records the published-source compilation and source hashes. The one retained module is runtime.wasm, SHA-256 eb2395e1c115980dd78dbb646a26f755a4119a2a1e086942e715240e07c34bc6. module-source-load-verification.json verifies all 126 observed source/module loads against the sealed inventory, across threads 0–2. Git is the source archive; no source tree, media copy or capture bundle was retained.

The unchanged registered dx_donuts has presentCap:60, its original EXE and 37 data files. All 38 hashes and sizes match the accepted November 7 pins before dispatch and after cleanup (accepted-media-match.json, media-before.json, media-dispatch.json, media-after.json). Launch uses normal guest Workers, real ticks, no-build and a finite 600-second CLI limit. The title and native Win98Renderer/raster Canvas2D output are DirectDraw evidence, not GL backend proof. No D3D corpus membership or historical27 completion is claimed. No fresh active-D3D counter was requested.

Production blood2-reviewed-input.cjs, the immutable preparation helper, sealed-source loader and common native transport were reused byte-for-byte from the published base. The controller permits the relevant ordinary Enter, arrows, Space and Escape gestures; its open-menu route name was used for the documented Escape exit attempt. No input helper semantics changed. Donuts-only orchestration is retained as run-local JavaScript and patches.

Before the original, the same sealed-source launcher ran two independently authored native USER windows. Actual full 1024×712 PNG construction passed on the main thread and in a Worker with the PNG codec included. The first native scope checked actual Right/Up/Space consumption, release child armed before DOWN, independently acknowledged UP, stale-review rejection and terminal no-further-endpoint behavior. The second deliberately discarded a real native-consumption ACK: independent UP still ACKed, the controller closed, and a later attempted capture made zero endpoint calls. Both scopes were cleaned and actually waited before the original. The whole authored binding desktop was also personally reviewed. These are input/preparation controls, not original-game movement/fire evidence. No tracked code changed; no runtime old-fail/new-pass or new locked-build claim is made.

Closure and recovery

Original PID2161408 ran from 01:15:03.083Z to 01:16:47.885Z. Its sole admitted input was the balanced Enter gesture. Host cancellation requested orderly shutdown; actual native exit0, both output EOFs and an empty drain-error list are retained. The independently armed finite/disk watchdog exited0 after seeing that actual native wait, with its own actual wait and both EOFs. native-supervise.cjs records pdeathsig containment, exit0, no deadline fire and leaked:false. Both authored scopes have equivalent native/watchdog/ containment receipts. All tracked native PIDs are absent or dead, and the control listener is closed. Sampled original free space stayed at or above 2,411,900,928 bytes; no RAM gate was used.

The board watcher was signalled with SIGTERM and its final absence/dead state checked. It had no retained parent wait owner, so its actual wait is explicitly unknown, not fabricated. The run retained seven selected whole PNGs, the unique module and authored PE inputs, receipts, JSON and logs. Redundant authored PNGs and duplicate module hardlinks were removed. Recover runtime source from the exact Git base; recover pngjs via the package-lock integrity; recover the unique compiled module from the retained run root. The run-local helpers and note are also preserved as new-file patches. result.json is written last after publishing this note.

No test box was created: box ID null, state NOT_CREATED, expiry null, stop receipt null. No agents/workers, Wine, Python, public deployment, foreign source changes or subsequent original replay were used. Ownership is released at task closure; no investigation remains held.