Forsaken pinned provider recovery and native Quit, 2026-11-14
The exact provider is recovered and tested: all162 reference sizes and SHA256
hashes match. The sole original reaches EXIT GAME and calls ExitProcess(13)
after260.339seconds, before the600-second ceiling. Native Quit is established;
this is not a clean guest exit0. Enemy damage/defeat and full27 remain unproved.
The final independent UP finds an already closed endpoint and is UNKNOWN;
the invoking tool returns143 and its additional outer wait is missing.
Evidence is in scratch/runs/20261114-forsaken-pinned-quit/.
Exact provider recovery
A local search under /home/user, /tmp, /opt and /usr/local finds none of
the nine missing native files. The search records permission errors explicitly;
it does not establish absence from inaccessible directories. Private
npm ci --offline --ignore-scripts --no-audit --no-fund restores153 matching
package files. It executes no postinstall and does not recover the native nine.
The predecessor's npm-only offline162 recovery description is incorrect:
the audited helper invokes curl independently of npm's offline flag.
The owner explicitly authorizes the five official default release archives. Each package's actual pinned install.js and addon-tools helper is audited and hash-pinned before fetch. Node24.18.1 selects uv-loop's Node24 archive. Default release tags differ from several npm package versions:
| Pinned package | Official release archive |
|---|---|
| deps-opengl8.1.1 | https://github.com/node-3d/deps-opengl/releases/download/8.0.0/linux.gz |
| glfw7.4.1 | https://github.com/node-3d/glfw/releases/download/7.3.1/linux.gz |
| segfault4.0.3 | https://github.com/node-3d/segfault/releases/download/4.0.0/linux.gz |
| uv-loop0.1.2 | https://github.com/node-3d/uv-loop/releases/download/0.1.0-24/linux.gz |
| webgl6.0.3 | https://github.com/node-3d/webgl/releases/download/6.0.1/linux.gz |
All five fetches succeed once through the inherited normal network/proxy
environment. No environment candidate override, proxy bypass, access-denial
retry, root install, media/source download or additional package download is
used. fetch-receipts.json records actual URLs, redirect URLs, archive SHA256,
byte counts and actual curl waits. Archive members are inspected before any
write: absolute/traversal paths, links, special files, duplicate names and
unexpected members are refused. Every extracted file matches the reference
size and SHA256. The complete153+9 installation is verified before loading.
The verified nine native files are retained as the requested ignored local
dependency cache at scratch/native-provider-cache/node3d-20261114/, outside
prunable runs. dependency-cache-index.json records their hashes and archive
provenance. This is a provider dependency cache, not a captured WASM/module dump.
The small pinned package metadata plus offline npm cache recover the153 files;
copying the nine cache files to their indexed relative paths and verifying
all162 reference entries completes future local recovery. npm's offline flag
alone is insufficient. Git contains no provider binary.
A second private installation proves that exact recipe using only the existing
offline npm cache and the preserved nine-file cache: all162 reference hashes
and sizes match, without network or loading the provider. The temporary second
installation is removed afterward. cache-recovery-check.json retains the
actual npm wait/EOF and final verification.
Retained predecessor
The original living-combat run's terminal files still match the six hashes
retained by the bounded-controller predecessor. Its late80 negative, terminal
input closure,5.745-second hold violation,900-second timeout and missing outer
wait are unchanged. It is never rebound, restarted or replayed. Five lives,
full hull/shield, outgoing pulses, two enemy craft and PULS1982 remain its
established living evidence; enemy damage and Quit remain unproved there.
At that predecessor's close, modal repair a56ebeec0 still required the actual
EXIT GAME route exercised below.
Checks and sole original
The launch waits for the current SuperTux original's actual terminal receipt,
then holds the heavy lock across preflight and the sole original. All29 durable
controller/lifecycle controls, three production guard/owner/transport suites,
the actual native headless GL test, driver identity and the full build under
/tmp/wine-build.lock pass. The controller suite retains its known late80 and
busy physical-bound negatives. No controller, supervisor, runtime or guard
source is changed by this worker. The native driver is Mesa llvmpipe,
LLVM20.1.2, WebGL1.0/native GL4.5, Node24.18.1, DISPLAY:0.
The build has zero drift across1647 source pins. Actual loaded WASM SHA256 is
c94ccf1b179f18bd4b47867445afed8cef2f1393b19aabfd259c7f3b94a7be09.
Git SHA plus binary patches identify dirty inputs; no source tree or module
binary is archived. The original uses direct production guarded HTTP8517,
synchronous native D3DIM, unchanged894 media/input pins and manifest/CRT,
real ticks/default1000. It has a fixed600-second ceiling and independent
prearmed620-second watchdog. There is one permanent owner, no remapping,
no guest-state/config/health/AI/physics/clock/blend writes and no restart.
Single Player / START Normal / VOLCANO is entered through ordinary menus. The level finishes loading and enters living play without a separate fourth Enter. Three initially stale combat intents are positively NOT_SENT, with no DOWN; the five-second scene bound is never overridden. A short run-local invocation wrapper reduces command latency without changing the durable controller. The successful manual A+Ctrl1800 action uses capture ages3.297 and3.357seconds and personal-review ages0.221 and0.281seconds. Its actual midpress shows five lives, full SHLD128/HULL128 and outgoing green fire. Postrelease shows forward progress into the corridor, continuing pulses, five lives, full hull/shield and PULS1988. Both edges have independent release ACK, actual release-child wait/EOF and postrelease progress; completion bounds are1.865 and1.805seconds from consumption.
The subsequent washed-out scene is death aftermath. An initial human reading
of its text as a kill counter was wrong and was explicitly withdrawn on the
messageboard. Reinspection and the original executable's enemy %s you /
killed strings identify “ENEMY KILLED YOU.” This is not enemy-defeat proof.
No additional aimed volley is sent using that scene. Target damage and kills
remain unproved; the living outgoing-fire evidence is retained separately.
Actual Quit and exact lifecycle limits
With more than200seconds remaining, ordinary Escape opens the native pause menu. Up selects QUIT TO MAIN MENU; Enter returns to the original main menu. Up selects EXIT GAME; Enter shows the original promotion; a final personally reviewed Enter dismisses it. Finite400ms menu gestures are justified by the repaired transport fixtures and this run's actual ACK latencies: maximum consumption-to-ACK2.771ms. All menu scenes meet30seconds, maximum15.030seconds. The audit contains10 consumed actions/11 edges and three positively unsent intents, one owner and no replay/rebind.
The trace records PostQuitMessage(0) at API9036672, then the original
MessageBoxA back-buffer error at9036682, then ExitProcess(13) at9036696
and [Exit] code=13. The modal with explicit pending quit does not remain
parked. This exercises the route previously missing for a56ebeec0; it
does not claim that the original back-buffer error has been repaired or that
the guest returns0. Native process3724096 actually waits0 with both output
EOFs; the harness's0 is distinct from the guest's13. Stats are260.339seconds,
with no max-seconds stop. The independent watchdog exits unfired because of
the actual native wait. Supervisor3720094 records native/watchdog actual waits,
releases the heavy lock and is itself actually waited0 by outer3720093, with
separate stdout/stderr EOFs.
The final promotion Enter is CONSUMED, but native termination completes within 0.239seconds of its DOWN, before the400ms independent release. The release child fires and gets connection refused; its actual wait is1 with both EOFs, and the controller permanently closes with UP UNKNOWN. No final physical-UP ACK is claimed. All nine earlier actions have release ACK within2seconds. The native is absent before the final release request; this does not turn its UNKNOWN release into ACK or authorize replay.
The additional invoking wrapper's tool-child-wait.json is missing. Its tool
session actually completes143, cause UNKNOWN. This preserves an outer lifecycle
gap despite the recorded successful native/watchdog/supervisor waits; there
is no claim that every required outer/tool wait and EOF passed. No second
original or further input is attempted to improve these records.
Postflight verifies zero drift in1647 sources,105 loaded files,894 media, 162 provider files and the loaded module. The nine-file cache remains; the private installed provider is removed after verification. Only selected checked PNGs, compact logs/JSON and Git patches remain in the run. No box, worker or deployment is created: box ID/expiry null, state NOT_CREATED, stop receipt not applicable. Full27 remains unproved.