SuperTux: native level entry, reviewed jumps and neutral landings

The one authorized original entered Welcome to Antarctica, stomped the first enemy, and completed 33 personally reviewed gameplay choices: 27 right-100, two jump-right-200, and four explicit airborne neutral-settle-200 choices. The fresh resulting images show native Pause, the jumps, score 50 after the stomp, and returns to the snowy floor. An elevated first-pillar landing, onward traversal, and the level exit remain unverified. This task did not achieve its full goal and did not establish a runtime defect that warrants a production change. No production code changed.

Evidence: scratch/runs/20261109-supertux-prompt-choice-gameplay-0229/. Selected exact reviewed images are level-visible.png, initial-pause.png, pillar-jump.png, settle03.png, pillar-jump02.png, and settle04.png. The complete captured frames, commands, input logs, original timestamps, module identities, reviews and failure output are retained locally and in the hash-addressed recovery set. No later game input or screenshot followed the first owned command failure.

The final successful settle produced capture review-3f89b244-6229-43f7-8da4-116dfa5f4f5c at 02:39:37.029Z, SHA-256 df380bd626a56149b4c095ea84848af096ab8bfc058897cecc2073b30e0d6552. I inspected it: Tux had returned to floor height in front of the pillar, rather than landing on its visible top. I deliberated too long over that geometry. The attempted onward01 helper began at 02:39:52.035Z, already 15,006 ms after the supporting capture, and failed at 02:39:52.142Z with review choice disabled. The page refused before click(), HTTP command submission, a numbered game command, or input. No frame was refreshed to evade this refusal. Only the ordinary no-input cleanup quit followed; it closes the browser and does not resume the game or dispatch keys.

choice-timing-audit.json.gz reads recovered choice metadata, not pixels. All 33 accepted choices were within the unchanged 12-second bound, at 4,941–9,135 ms. Twenty-eight met the 8-second aim; five did not. The failed following choice was outside 12 seconds. Each resulting native image was personally displayed and reviewed before the next selected action; there was no diagnostic/tool roundtrip between image review and that choice. Every airborne review used only empty keys and fixed 200 ms neutral settle, followed by a fresh native Pause review. There was no route, detector, replay, retimestamp, limit extension, or freshness workaround.

After the run stopped, a static check of the original level file identified a material geometry distinction. At world columns 24–25 (x768–831), the first visible snowy pillar uses tile IDs 122/123 in background row 11 and 124/125 below it. The corresponding interactive-layer cells above floor row 13 are zero; floor tiles are 8. The second early pillar at columns 28–29 has the same layer placement. The tile definitions name snowbg1.png through snowbg4.png and include solid #t, so the layer distinction matters; tile flags alone do not prove a collision surface. This supports a scenery interpretation of the apparent pillars, consistent with the observed descent to the floor. It does not establish an emulator collision failure. static-pillar-layer-check.json.gz records the exact original file paths, sizes, SHA-256 values and relevant cells without copying media or generating gameplay decisions. A future task should distinguish the requested scenery checkpoint from a real interactive landing surface before claiming success.

Offline preparation read the 0212/0124/0116/0046 notes and included published df0932c1c terminal-failure replies. The run's canonical Git was 9944ddf9227cc5062ed0604b45687bbc3826bb99. Local/remote barriers matched 1,462 providers / 61,453,389 bytes, manifest SHA-256 a865214924efc8fe75624f942c6720a341a3a8fae27b12105376d513db3a7d0a. The loaded CPU SHA-256 stayed a320d4c85156644fd12b6134d66b11784f3100b9b135756206c087cdacd43295; native observation stayed OFF. The exact singular physical-key Enter, ArrowDown and Escape payloads passed the actual composed executor offline before capture. Accepted unchanged controller/client/composition controls and the 0212 focused failure-reply test and locked build were verified and reused by PATH+HASH. No new build or unchanged-test rerun is claimed. Preparation initially referenced a missing historical test path, then detected a stale provider fingerprint after that correction; both failed checks and the corrected final barrier are retained. Both happened before allocation.

The exact Warcraft0224 bx_gsub82ym ARCHIVED PHYSICAL_DRAINED release was posted at 02:32:00Z; the interruptible watcher actually waited 0 and awaited its tail. No peer box was accessed. The independent immutable watchdog was prearmed at 02:33:00Z, UUID cf9c2687-ab72-4115-ba9c-e30d8388a6ae, PID 865281, start 223878940, deadline 02:58:00.114Z. The sole fresh box bx_3zcmeteb was created at 02:33:08.903Z with TTL 1,491 seconds and initial archiveAfter=02:57:59.745Z; expiresAt was not returned. The original began at 02:33:44.490Z. There was no second original.

Cleanup Browser.close received ACK and completed normally in 207 ms, browser exit 0, signal null, no escalation, with terminal streams drained. The client exited 0 with separate stdout/stderr EOF. Custody verified 373 owned identities / 373 actual reaps / 3,960 events, including the before-wait PID/start witnesses. Kernel parent 6303/start8530 and outer parent 6300/start8519 actually waited 0 with separate stdout/stderr EOF. Ports 38843, 41041 and 8787, writable run FDs, profile, input/client locks and runtime lock were drained. Both physical-drain records show pending 0. The watchdog actually waited 0 and its tail -15, with stream EOF.

The local runtime and watchdog-supervision tool terminals returned 143. Their wrapper exit receipts and the local Boat launch's actual-wait receipt are missing and remain UNKNOWN, even though the remote waits are complete and all locally recorded PID/start identities were later absent. The release tail's start identity was not recorded. One truncated local stdout gzip and five empty local gzip paths are retained verbatim with SHA/size and UNKNOWN CRC in local-gzip-audit.json.gz; no footer or wait was fabricated. Historical untraced-clone, nonleader-exec and external-signal limits remain UNKNOWN.

All 662 remote regular files were accounted for, including pinned input references. 544 unique transports / 13,375,920 stored bytes passed original and stored SHA-256, size, independent reopen, GNU gzip CRC and exact-set checks before delegated normal stop. The box was archived at 02:41:43.213Z; receipt stop_64e3904c896da4ae completed at 02:41:43.231Z, forced=false, error=null, final archiveAfter=null, expiresAt not returned. Physical release was posted at 02:42:22Z. Free /home stayed above 1.5 GB at recorded checks, with narrow headroom reported to the coordinator. Git SHA plus small owned input patches preserve source recovery. No source/media/module archive, peer cleanup, local browser, local heavy execution, parallel runtime, agents, RAM gate or deployment was used. Next work requires a separately authorized original, an accurate landing checkpoint, and decisions submitted before the unchanged freshness deadline.