Alice observed timeout — 2026-11-04

The original batch1681 ProcessDesktop timeout remains unexplained. One changed diagnostic completed the formerly failing1666→1702 request and reached ordinary native ground movement without reproducing the timeout. There is no causal repair, scheduling-cause attribution, completed gnome dialogue, knife acquisition or gameplay-completion claim. The original timed-out destructive poll's outcome remains unknown; it was never retried or acknowledged.

Run: scratch/runs/20261104-alice-observed-timeout/. Prior evidence: original failure and participant observations.

Exact inputs and authorization

Baseline 625be68f9c2b685797a81f837992f4421d469f16 includes observation 458625894d20887053d23227152abb5467bf78ae. All385 tracked source/font/compile inputs were read with git show SHA:path; foreign WIP and subsequent main changes were excluded. source-identity.json and empty relevant-tracked.patch identify that closure. Git is the source archive. owned-helpers.patch preserves the small run helpers; postlaunch-recovery-helper.patch preserves the later local-only recovery helper. No source/media archive was made.

All43 original dependency/media pins and the original manifest matched before launch and after execution. Media was transferred directly into the runtime, without copying it into evidence. The baseline uses accepted458625894 locked-build modules; its compile-source paths have no committed difference from625be68f9. Both were recovered by full size/SHA256:

These differ from the original failure's065267ce/e7c59d3d modules, retained by prior path/full-SHA references. This is an explicit diagnostic input difference, not a controlled replay of the old modules. baseline-module-provenance.json links the accepted build. PID8141's observer proves canonical compilation and instantiation; compatibility was retained, not executed. Observer filenames include PID/Node threadId to prevent worker preloads overwriting the main record.

Preparation used syntax and small fixtures only, with no box/build/browser/game. Coordinator board01:18:55 explicitly released Alice after MDK2's verified normal STOP01:18:38.420Z. This worker acknowledged before creating fresh bx_u3jnuyts, TTL3600, expiry 2026-11-04T02:19:09.797Z, shutdownOwner=coordinator. Identity and expiry were recorded before use. No predecessor was resumed or worker spawned.

External observation

The Python supervisor gates child execution on a pipe until its baseline is saved. Thus external sampling began before Node PID8141 could create workers. Sampling ran every nominal50ms with128 retained samples (nominal6.4s; actual monotonic start/end and wall times expose gaps). Recent, prefix-checkpoint and final windows retain task appearance/disappearance, stat/status/affinity/wchan, schedstat/sched, process CPU, host CPU/load/pressure, resolved cgroup membership and the mountinfo baseline. Per-thread on-CPU/run-queue/timeslice deltas and host/process CPU deltas accompany raw files. Cgroup-v2 quota/statistics/pressure/ cpuset, v1 equivalents and visible ancestors include interval counter deltas. Read failures retain errno. Task sampling caps at64, ancestor traversal at16, history at128 samples and first-error trace capture at32 files.

Nine OS threads were observed, including all worker-written identities: main TID8141, authority TID8148, bridge TID8149, Node v24.22.1. Checksummed participant records retain correlated request phases and bounded preload identities. Frozen prefix/final summaries contain no recorded first error or timeout. On cleanup, retire1857868 completed before termination callbacks reported code1 for both workers. Those close-time callbacks do not establish an original worker failure.

last-successful-poll.json correlates poll1857867 across main submission, bridge/socket/authority receipt, handler completion, reply-write, bridge frame, shared notification and main return. Limit256/local pending0, queue0→0, returned0. There is no failed request phase or timeout response-word snapshot in this run. The synthetic fixture exercises first-error capture; it is not an Alice failure. The3000ms deadline, fail-closed/no-retry and unknown-outcome handling were unchanged.

The sampler ran continuously through retirement, but rings are recent history, not a complete audit. No actual3000ms failed wait occurred. Synchronous tracing can perturb timing. perf_event_paranoid=4 restricts scheduler event tracing; samples cannot recover every short transition. Missing JS markers, absent error records or aggregate CPU do not establish an OS scheduling cause. None is claimed.

Native progress and bound

The normal local-entry route used software rendering, cooperative --no-threads, frozen scheduling, batch size100000, --trace-input, canonical Wasm and --max-seconds=900. No held-phase adapter, guest writes/injected calls, configuration/game patch or cheat was used. Original payloads were segmented for capture review without replay. The final small segment's SHA was recorded before use. All124 sequential requests completed:103 step requests, max64, total4926 batches. Fifteen native events contain seven balanced key down/up pairs and one relative mouse movement; no input remained held.

Personally reviewed captures:

Last gameplay command completed at847.401754s, within the15-minute segment. The worker missed its early-quit checkpoint. The frozen idle runner did not retire at its900s option; with no further gameplay input, an external ordinary-stop caused supervised SIGTERM/CLI cleanup and exit0 at917.637017s. Process lifetime therefore exceeded900s by17.637s, with70.235s idle after the final command. This was not a normal control-quit receipt or a strict900s process-bound pass. No pending command was retried.

The original IPC failure is the unresolved causal blocker: this changed run did not reproduce it, so there is no tested repair. Native acceptance is limited to the reviewed ground route. Planned dialogue waits do not prove that a dialogue trigger or knife objective completed. No unchanged replay followed.

Recovery and handoff

Recovered667 unique outputs,24836058 bytes, individually by full SHA256, plus both modules. The exact remote output set was stable before/after recovery. All458 sealed inputs and the provisioning manifest's own SHA remained unchanged. Source/run inputs and media were not re-archived; the original-log symlink was recorded without duplicating its target. recovery-verification.json maps all remote outputs locally, including two raw launch records kept separately from their differently structured local API-wrapper records.

Checks passed: helper Python/JS syntax; cgroup-v1/v2/subtree and short Python process fixtures; launch gate/live task/delta collection/bounded ring/synthetic first-error freeze; receipt interface/forbidden-command/large-step rejection; 124 terminal receipts/native input balance; canonical compile/instance identity; full input/module hashes and stable exact output-set recovery. There was no production code change, causal regression or build by this worker. The accepted module build is explicitly referenced above; fixture success is not gameplay.

At handoff, bx_u3jnuyts was RUNNING/idle, expiry 2026-11-04T02:19:09.797Z, shutdown owner coordinator. Original PID8141 and control8123 were absent/closed and driver output stable. The worker did not stop, delete or resume the box. Coordinator verification/snapshot/normal STOP is next. Publication details are retained in publication.json in the run. Single task ends here, without a successor or deployment.