SuperTux native confirmation contract and platform landing, 2026-11-18
Manual native play reached and settled on the first raised block platform in
Welcome to Antarctica. Ordinary Quit remains unachieved: the next Escape was
refused before DOWN because the controller queried a compound aggregate nonce
that guarded-status does not retain. This task changes no production code.
Both new runtime scopes are permanently closed; host cancellation is not Quit.
Evidence is sealed in
scratch/runs/20261118-supertux-native-poll-confirmation-contract-course/.
causal-findings.json, individually compressed passive observations, public
input traces, original receipts and selected whole frames distinguish measured
facts from unresolved causes. Existing UNKNOWN scopes were never contacted.
Native poll delay: narrowed, not repaired
The first fresh scope used the same coherent 253-source graph and module
0d383140f6d4207d18860e3cb22fc8fcc47a5d260b8d5f2443222396b245fdd5.
The original 789 media/settings pins passed preflight. The existing sealed
module was reused without compilation. Main/thread quantum remained
10000/1000 from the prior incomplete cadence scope; neither guest time nor
poll/state/consumption was forced.
Title Enter passed actual native USER consumption at about 400 ms. Slot Enter then became UNKNOWN. Relative to public prepare, the handler entered at 17.705 ms and returned at 18.369 ms; delivery became pending at 18.590 ms. Transport disconnection cancelled it UNKNOWN at 492.581 ms. There was no native dequeue or USER-return receipt for that DOWN. Independent UP completed about 1197 ms after dispatch, within the unchanged 1750 ms bound.
During delivery-pending through cancellation, 124 observed main slices used
4.990 ms wall time and 4.854 ms process CPU. In 122 of them, both entry and exit
were yield reason 1 at EIP 0x5eef19. Across the entire observed window,
330/334 slices had that wait state (16.866 ms wall, 10.125 ms process CPU).
No sampled event-loop gap exceeded 30 ms during the pending interval.
The original loaded SDL.dll base is 0x5ca000; this instruction maps to
original 0x10024f19 in exported SDL_mutexP. Disassembly shows an infinite
WaitForSingleObject call immediately following it. Thus this new failure is
predominantly repeated admission of a waiting owner, rather than sustained
guest computation or guest sleep. A scheduler label saying runnable records
admission, not forward progress. CPU samples include Node work and are not
exclusive guest instruction accounting.
The second fresh scope additionally observed public mutex create/wait/release calls without changing arguments/results. It did not reproduce the slot failure: Start, slot and story Enter passed. Its bounded mutex log reached 16384 rows and dropped later records, so it does not establish continuous mutex ownership at the later jump. In particular it cannot prove a bad mutex owner, a missing wakeup, or the reason for the predecessor's 466 ms queue wait. A legitimate slow consumer is not itself a bug.
Response budget and ticket lifetime
The controller's HTTP confirmation limit is 500 ms. A ticket's immutable deadline instead comes from targeting freshness, operation completion and the release cap. Transport disconnection marks a pending ticket UNKNOWN. These are distinct contracts, not proof that a late native consumer is defective.
The new failed slot's handler had already applied input about 18 ms after prepare. It therefore cannot honestly be classified definitely unapplied. Early cancellation could only establish no effect while still queued before handler entry, using server-owned state. No caller no-effect flag, receipt substitution, timeout increase, replay, or cancellation repair was introduced. The published 09d56288b release reserve remains 1250+500 within 1750 ms.
Manual landing and the exact next failure
Whole frames were personally inspected before finite manual gestures. Stale overworld reviews were refused locally without DOWN. For the static map, one new capture was required to match the exact PNG hash of the personally reviewed whole frame before issuing the individually chosen Down or Enter. A mismatch would refuse that gesture; there was no automatic retry or action selection.
After entering the course, Right for 550 ms and Right+Space for 600 ms both
received native USER-return receipts and independent UP acknowledgements.
mutex/c2.png shows the initial ground position, c3.png the approach,
c4.png the raised-platform position, and landed.png the settled landing.
The settled frame shows score 0, coins 0 and four lives. The camera scrolled
right; this is platform/course progress, not score gain or course completion.
The next manually chosen Escape failed in lease(), before arming or dispatch:
unknown guarded nonce. This is a concrete compound-status contract mismatch,
not ticket retention expiry. compound() reserves its aggregate nonce in a
Set, while only its /0 and /1 edge tickets enter the receipts Map.
status() consults that Map. The controller saves the aggregate nonce in
last-intent.json, then queries it on its next pulse. The preceding successful
native chord receipt proves both edge nonces consumed; the absence of
pause-dispatch.json and the terminal failure identify the refusal boundary.
This finding is not published as a repaired input contract. A future change must validate subsequent lease/ordinary Quit after a successful nonempty native chord, plus late consumer, partial effect, lost/late reply, refusal and independent release controls, run the locked build, and exercise the affected original in that same task. It must preserve every existing UNKNOWN closure.
Lifecycle and recovery
Finite detached native-supervise wrappers, individual native/watchdog/holder and parent waits with EOF are retained. Browser and display were not created. Both native waits exited 0 after explicit host cancellation; neither is Quit. The second outer tool invocation returned 143 despite nested parent/native waits returning 0; that invoker limitation is retained rather than relabelled. All 24 recorded owned PIDs were absent at the final read-only check.
The task initially encountered free disk below 1.5 GB and launched no runtime until other owners restored capacity. Each runtime sampled the floor every 500 ms; this is not continuous global capacity proof. All test growth was owned tmpfs. Selected whole PNGs and unique nonmedia evidence were retrieved, fsynced and hash/roundtrip verified before owned tmpfs removal. Runtime Git identities, binary dirty patches, source pins, owned new-file patches and the module recovery recipe are retained; no source-tree or media bundle was copied. No worker, Wine, remote box or deployment was used. Box ID/state/expiry/stop receipt are null/NOT_CREATED/null/null.