SuperTux controller composition and one visible attempt, 2026-11-08
One fresh original reached Welcome to Antarctica, collected block coins and visibly flattened the first ground enemy. The final reviewed native Pause boundary shows score 175, five coins, four lives, timer 276, with Tux airborne above the next ground enemy. No death, level exit, unlock or next-node change was observed. The attempt stopped on the first controller UNKNOWN and was not retried. Historical 675-point/23-coin progress is separate route guidance.
Evidence: scratch/runs/20261108-supertux-controller-visible-progress-2301/.
result.json.gz, control-summary.json.gz, remote-inventory.json.gz and
retrieval-verified.json.gz identify the original and its retained artifacts.
review/terminal-boundary.png is the final personally reviewed image.
The Q3 worker's explicit 23:01:29 RELEASE preceded allocation. Its retained
archived state/stop receipt and actual waits, closed ports, absent writers,
retired profile and drained locks were independently checked by PATH+HASH.
The only new box was bx_uwk628ca, returned with TTL 1495 seconds and
archiveAfter=2026-11-08T23:31:24.127Z. The service returned no expiresAt
field. A durable UUID and observable independent immutable 25-minute watchdog
were armed before its sole create. No local Chrome, second runtime, agent,
old-box access, resume, delete or public deployment was used.
Exact composition and observed failure
The offline check verified all 1,361 non-scratch retained runtime providers
through their recorded Git blobs or existing media/module PATH+HASH. The current
working files were also compared: Q2's working guest modules were different and
were not copied. The remote final startup barrier independently checked
1,383 providers / 61,300,499 bytes, then its consumer checked the same
generation immediately before launch. The source graph retains the prior
published runtime selection 0e2bd19d802278aacc06dd952e135529c1edad63, with
explicit published helper replacements. The actual executing module was
4b45bda53a90b665896cf3b85d9a157b7c3703563c88b6666e4ee2e4b346bcc3.
The accepted integration patch was applied in memory to the exact retained
browser-fixed.cjs, SHA
9f87baf0b9a22ae3caff7ae3072a6ce89fc270ca107dfe9653b18f882f18292b.
The controller remained published d3f172fee, the actual readiness/terminal
consumer 80145daf, and queued-event custody dec33253c. The new composition
helper was published in 58c56eb7b before staging. The final post-run repairs
described below were tested offline and were not rerun on an original.
No source archive, media copy or executing-module copy was retained; Git SHAs,
small owned patches and existing provider paths supply recovery.
The browser started at 23:08:32.205Z, reached its genuine native menu, and followed the ordinary slot/story/map route. The two initial Pause reviews show timer 287 unchanged. Twenty admitted compositions completed 60 phase pulses / 134 actual controller CDP operations. The native event recorder contains 76 trusted DOWN and 76 UP, with no held key at closure. Each phase's previous actual completion was checked before the next phase's DOWN. Ground contact was personally reviewed before jumping; actions near enemies used separate 100–200 ms phases. The first enemy was visibly flattened with score increasing 100 → 150. Native Pause was re-reviewed after accepted actions. Pause transport completion alone was never treated as grounding or survival.
The capture enemy-two-jump began at 23:15:40.357Z. The proposed next
100 ms Right action reached admission at 23:15:52.770Z, 12,413 ms later.
It was refused before dispatch with zero effects, so that action did not
run. The persistent controller also recorded sticky UNKNOWN/noReplay.
The generic retained transport nevertheless returned its refreshable-stale ACK.
The orchestration then failed to retrieve an action screenshot that did not
exist. This is the precise new composition failure; there is no claim that the
airborne enemy transition was completed. Only a read-only terminal capture and
ordinary runtime quit followed. No further gameplay DOWN or second life ran.
tools/supertux-native-driver-composition.cjs provides strict, hash-checked,
in-memory patch composition and retains actual input CDP promises until reply
or explicit disconnection. An ordinary 20-second CDP timer can no longer discard
those original input promises. Shutdown cancels the controller and awaits its
pending promises and physical helper before closing logs. The post-run repair
preserves controller UNKNOWN across the transport boundary instead of allowing
the generic stale-refresh ACK. It also corrects the accepted persistent save
closure's first-command ID on subsequent phase rows. Original evidence is
unchanged: those phase IDs are reconstructed from admissions and global command
numbers, with that limitation recorded in control-summary.json.gz.
Focused controls exercised the actual retained gzip command consumer, published
controller and lifecycle consumer. An actual Node child delivered independent
UP while DOWN remained pending past the ordinary CDP timeout, then its genuine
late DOWN reply. A pending close-request promise did not substitute for awaited
child terminal status and inherited stdio EOF. The stale-controller regression
checks zero input effects, a terminal failure artifact and no refreshable ACK.
Hash mutation, missing lifecycle provider and malformed patch hunks are refused.
All fixture children were awaited and finite guards cleared. Final focused
controls and flock /tmp/wine-build.lock bash tools/build.sh passed.
The run's small JSON storage adapter preserved the exact retained helpers and atomic filenames while compressing driver receipts/markers before writing. Its first control exposed double compression of an already-gzipped partial command; that failure is retained. The corrected actual-consumer control passed.
Drain, recovery and normal box stop
The browser's real close receipt is normal=true, CDP close ACK, exit 0, signal null, no escalation, terminal and stdio drained in 11,704 ms. Pending physical input was zero. The actual kernel journal contains 3,731 events / 352 registered PID/start identities / 352 actual reaps; before-wait witnesses were independently checked. Both parent terminal waits returned 0 with stdio EOF. Ports 45361/32885, own writable FDs, input lock, active profile and runtime lock were drained. CLONE_UNTRACED, unsupported nonleader exec and external-signal limitations remain explicit. An empty process list was not used as a terminal-wait substitute.
All 384 remote regular files were inventoried without extension filters:
four exact source references and 380 evidence records, covered by 304 unique
gzip transports / 10,456,154 stored bytes. Every transport was independently
reopened and checked against stored/original SHA-256 and size; gzip CRCs,
including nested gzip artifacts, were checked before normal stop. The new
box was promptly stopped after recovery: archived, stopped at
23:18:23.081Z, archiveAfter=null, no expiresAt field returned. Receipt
stop_9a75b94225559e53 completed 23:18:23.100Z, forced=false, error=null.
The independent watchdog and its parent were awaited. The bounded board watch
used a same-shell EXIT trap; tail actual wait 143, gzip wait 0, deadline wait 0,
and the outer shell/command completed. Local free-space samples stayed above
1.5 billion bytes; no continuous-floor or RAM-gate claim is made.
Evidence limitations: initial read-only inspection tool outputs preceded the
gzip command wrapper and were not retained as run command logs. The unchanged
kernel writes its two JSON metadata receipts raw before their gzip recovery,
and an initial post-run stale fixture wrote raw receipt/marker artifacts.
Those exceptions are retained and not claimed to meet gzip-first storage.
Driver commands, streaming logs and the corrected storage path were gzip-first.
The final audit also found that the initial board watch and its command wrapper
both opened boardwatch.log.gz, corrupting that local log. Its stored bytes and
SHA are preserved; decoded original SHA/size are UNKNOWN and its CRC check
explicitly fails. The separate Q3 release/state/drain checks and all downloaded
remote payloads verified independently. The watcher filename was corrected and
a final five-second watch reached awaited tail/gzip/shell termination with a
valid gzip stream. This does not reconstruct the damaged initial watch output.
The final local inventory covers every regular file, including those exceptions
and owned source inputs, with independent stored hashes/sizes and decoded
original hashes/sizes wherever decoding succeeds. It explicitly records this
one damaged-log exception rather than claiming an all-file CRC pass.
Next work requires separate authorization for a fresh original and should use the corrected UNKNOWN propagation with a newly captured/reviewed boundary that fits the unchanged 12-second admission window. This task stops here. Full27, parity, audio, FPS and level completion remain open.