NFS3 prepared manual workflow after loader repair, 2026-11-18

Blocked before the original: free home space crossed the 1.5 GB floor during the authored manual fixture. Its scope is permanently UNKNOWN_CLOSED; there were zero original launches. Original driving, steering, native Pause, Continue and ordinary Quit remain unverified.

Evidence: scratch/runs/20261118-nfs3-driving-pause-after-loader-repair/. The existing published startup remedy 0888e0e4c remains applicable; this task made no tracked runtime changes and did not repeat broad qualification.

The async-owner 9191d4afb proof is reused: native event turn 2642.46 to 0.876 ms, equal pixels/order/retirement, sequence 73, arena tail 12292, twelve focused suites and locked build. The loader proof is also reused: old native hello ETIMEDOUT, repaired hello/snapshot 664 ms, authentication/cancel/error, inherited Worker source inventory without Worker Git and locked build. The earlier 32-draw consumption UNKNOWN stays closed; responsive host execution does not guarantee an application's USER polling interval. The later 4210-draw, 131-present fixture's first capture missed its normal 105-second exit; that was an operator missed window, not a polling defect. Neither endpoint was reused.

Preparation and the new finite scope

Before launch, preparation verified published source 0888e0e4c6246d8782690d8ca437d708038ad5bd, the repaired loader and passive observer against Git, retained module 7d347a6a631ab77335a14935a533feb896be71aa2508fc5aa19b1e5df8f096ce, the unchanged genuine chooser registry and all 135 media pins. Launcher, control, independent-UP, watchdog, original command and local readbacks were prepared. The original command remained conditional on a complete manual gate.

The authored positive control derives from the existing healthy one-draw Device2 RGB565/D16 fixture. Its draw count and USER/Sleep loop are unchanged. The only added guest branch handles Q with a log, 1200 ms cleanup allowance and ExitProcess(0); that branch was never exercised. It is not a modified 32-draw fixture or a claim that the earlier 32-draw scope can pass. Source, executable, base hashes and regeneration recipe are retained; no media was copied.

The scope used the same pinned loader/async owner/public CLI route, software cooperative renderer, independent native-supervise containment and finite/disk watchdog. Its native limit was 180 seconds; it was never extended or reopened. Main and four native Workers record the correct source/module identity and inherited Worker inventory. Initial free home space was about 2.39 GB.

Observations and limits

The first whole 800x600 canvas capture arrived about 15.6 seconds after native startup, while live. The automatic capture trigger waited for a suppressed debug string, so this first capture was issued manually. The selected images show the authored Device2, PAUSED and CONTINUED captions on a grey GDI canvas. They were personally inspected. They do not establish the contents of the software D3DIM backbuffer, a composed triangle, or any original-game frame.

After personal review of the initial frame, separate Up, Left, Right and Escape presses confirmed native owner USER-return consumption in 87, 86, 250 and 249 ms. Host application preceded native confirmation by 10, 10, 15 and 33 ms respectively. Every press had an independently armed UP at the unchanged 800 ms deadline, successful release acknowledgement and actual helper wait/EOF. These are native delivery receipts, not authored acceleration/steering gameplay consequences or an engine polling guarantee.

One Continue receipt confirmed native consumption in 82 ms and changed the caption, but its command was issued in the same tool call as image inspection, before the image result was available. Its saved reviewed:true is explicitly excluded as proof of prior personal review. After separately inspecting the CONTINUED frame, a new Pause confirmed in 292 ms; its PAUSED frame was then separately inspected. The following Continue request failed ECONNREFUSED after the watchdog had stopped native execution. It permanently closed UNKNOWN. Its prearmed independent-UP helper also failed to connect; its actual exit 1 and separate EOFs are retained. There were no later non-release endpoint requests, retries, replay, scope extensions or original launch. The complete reviewed Continue/Quit workflow was not achieved.

Resource stop and cleanup

The watchdog sampled 1,490,628,608 bytes free, below the unchanged 1,500,000,000-byte floor. Native supervision records supervisor SIGTERM, guest exit 0, actual supervisor wait, separate stdout/stderr EOFs and leaked:false, ending at 01:28:11 UTC. This is host/watchdog termination, not ordinary native Quit or an application polling failure.

The inherited run watchdog has a receipt defect: its DONE handler overwrote the earlier stop reason with actual-native-wait/fired:false, and its pending shutdown timer subsequently raised ERR_IPC_DISCONNECTED. Its actual wait is exit 1. Raw receipts are preserved. Disk-triggered termination is inferred from the below-floor sample, watchdog code path and supervisor SIGTERM timing; the overwritten raw reason is not presented as direct stop-reason proof.

All twelve known launcher, supervisor, watchdog, native and release-helper PIDs are absent. The roughly 0.5 MB owned run cannot recover the shared disk deficit; no foreign files were removed. A local terminal-refusal test confirms zero fetch calls after UNKNOWN. Media hashes remain unchanged. No deadlines, polling, confirmation predicates or runtime sources were weakened. No Wine, Python, boxes, workers, deployment, archives or capture bundles were used. Box ID null, state NOT_CREATED, expiry null, stop receipt null. Evidence is finalized with result.json last, then ownership released and work stopped.