SuperTux short-step attempt: startup refusal and lifecycle fix

One fresh remote original failed before browser readiness. No game input, native Pause review, same-life progress, death or exit was observed. There was no second browser or original retry. The concrete failure is the retained SuperTux driver's fixed ten-second DevToolsActivePort polling budget.

Evidence: scratch/runs/20261108-supertux-short-step-progression-2236/. The prior native Pause result, published in cabc64e3b, remains separate evidence. All 4,533 retained remote references were independently reread against their original SHA/size, decoding the subsequently compressed files where necessary. No prior captures or qualification runs were repeated or copied into this run.

Actual original and failure

The staged browser/runtime source was published 0e2bd19d802278aacc06dd952e135529c1edad63, with individually pinned existing media/module paths and owned driver patches. The staged canonical CPU SHA was 4b45bda53a90b665896cf3b85d9a157b7c3703563c88b6666e4ee2e4b346bcc3; there is no executing game-module claim because readiness was never reached. The tested custody kernel was published bcf8c7bd788f2c63668406b11f72652006cde8dc, SHA 5371233b12d3a63e13a9768529af43e5a61b7ff1682b2d3c09e4db6a976e12d0. The concurrent parent-lineage worker's edits were neither staged nor absorbed.

The actual browser execution window began 22:37:30.533Z. At 22:37:40.672Z, the driver read the still-unavailable DevToolsActivePort after its 100 × 100 ms loop and recorded ENOENT with noReplay=true. Retained Chrome stderr announces an endpoint on port 40365, but that line has no host timestamp; its exact publication time is unknown. There was no CDP connection, accepted action or screenshot. An earlier menu capture request failed before its writer lock could be created and dispatched no input.

The failure cleanup escalated: Chrome's actual exit was SIGKILL at 22:37:57.189Z, rather than normal Browser.close. The input completion set was empty. Kernel journal, outer-launch and runtime-parent terminal waits, including their stdio EOF, are retained. The initial detached wrapper had already failed before any runtime because the final staging barrier receipt was missing. Its stderr was discarded by that wrapper; this evidence gap is explicit. The missing receipt was subsequently diagnosed without a runtime, published through the accepted barrier consumer, and the sole actual runtime was launched with a parent that waited and retained its compressed stream.

Concrete driver changes and checks

tools/supertux-native-lifecycle.cjs now composes the accepted native-devtools-ready.cjs and native-browser-close.cjs consumers for this driver. Readiness waits for a complete endpoint within the lesser of the original runtime deadline and 120 seconds. It never launches or retries Chrome. Cleanup requires terminal status and stdio drain before profile removal or quit acknowledgement. The corrected retained browser-fixed.cjs calls this helper at both boundaries; it was syntax checked, not rerun on an original.

Focused controls use three actual Node child fixtures, with an independent finite cap armed before the first child. A complete endpoint published after 11 seconds succeeds; a partial endpoint fails at its unchanged short deadline; exited/expired children are refused. A pending close-request promise does not prevent actual child terminal/stdio completion. All children were waited. These are real accepted-consumer controls, not native game claims. The required locked build passed after the helper and test were final.

The run-specific input driver also tightens capture freshness from 120 seconds to 12 seconds, retains original action/release timestamps, and records real DOWN/UP promise completions separately from bounded acknowledgements. Its focused controls passed with pending DOWN, independently prearmed UP, late completion and sticky UNKNOWN/noReplay. The actual transport consumer accepted gzip commands and refused subsequent input after an UNKNOWN fixture. These controls used no Chrome or game. The intended landing/enemy strategy remains unexecuted: first visibly review landing/ground contact, then use separately reviewed 100–200 ms movement/jump phases and inspect the actual native Pause menu and timer after every accepted active action. An airborne Space pulse is not evidence of a grounded jump.

Retrieval, custody and normal box stop

A durable UUID and observable independent immutable 25-minute watchdog preceded the sole create. Box bx_nn42bp5z was created at 22:35:13.268Z with returned archiveAfter=2026-11-08T23:00:08.128Z, ttlSeconds=1495. The service returned no expiresAt field; none is invented.

This actual kernel journal contains 1,297 events, 127 registered owned PID/start identities and 127 actual reaps, with their before-reap witnesses independently checked. Its recorded complete=true, empty remaining/unreaped and empty unresolved-lineage sets describe this run only. Historical unresolved clone-parent attribution remains unresolved. CLONE_UNTRACED and nonleader-exec limitations are retained; an empty process list is not substituted for waits.

Both ports 46525/40365 were closed, the runtime lock was released, input writer lock absent, and no owned writable FD remained. Chrome recreated a profile during the early cleanup race; after actual waits and writer drain, that directory was retired as evidence and the active profile path was absent. Its unique files were recovered, not silently deleted.

All 35 remote regular artifacts, 561,573 original bytes, were covered by 34 unique gzip payloads, 113,668 stored bytes; selected immutable source inputs remain referenced by the pin manifest. Every payload was independently reread against stored and original SHA/size, including nested gzip CRC checks, before normal stop. Remote symlinks/special files were enumerated: zero. Runtime and Chrome streams were compressed from their first stored byte. Early orchestration command-attempt metadata was initially raw JSON and later losslessly compressed; it is not claimed to meet gzip-first storage.

The box was normally stopped at 22:39:29.490Z. Final state archived, archiveAfter=null; receipt stop_0d30a65543c5964a completed at 22:39:29.518Z, forced=false, error=null. This normal box stop is distinct from Chrome's forced failure cleanup. Normal/watchdog/supervisor actual waits all completed. Final local inventory covers every regular file and records symlinks without following them; own small patches and Git/provider references provide source recovery without tree/media archives.

No old-box resume/delete, public deployment, other runtime, local Chrome, agent, RAM gate, guest/config/backend/seed/binding repair or prior-run deletion was used. Local free-space samples stayed above the 1.5 GB resource floor; the new evidence is small. Full27, parity, audio and FPS remain open.

Next: a separately authorized fresh original should pin the reviewed lifecycle adapter and refreshed published kernel, require its final staging barrier, then attempt the landing/enemy strategy. This task stops after its one failed original and concrete driver repair.