SuperTux compound key context, 2026-11-18

Evidence: scratch/runs/20261118-supertux-compound-key-context-course-progress/. Code: d6c398cf6 (strict pull/push succeeded).

The predecessor scope remains permanently closed. Its run-local pair helper sent Right and Space as two independently nonced guarded-input requests, each carrying the same all-keys-up snapshot. controlSnapshot().physicalUp is false whenever any renderer physical key or mouse button is held. nativeScene.context() includes that value; contextFailures() both requires physical UP and compares the complete context. A consumed physical Right therefore suffices to refuse Space. This is an input-contract/helper mismatch, not a changing image hash. The old original did not retain every post-Right context operand, so additional changes in that particular original remain unknown; no endpoint was contacted to reconstruct them.

An actual animated SDL/HTTP fixture reproduces the isolated cause: the only context operand difference is physicalUp: true -> false; its failures are input.physicalUp and context.unchanged. Right has an actual native USER return receipt and Space is cancelled before its physical handler.

The new guarded-key-chord endpoint accepts a bounded, declared sequence of distinct nonmodifier DOWNs under one nonce and immutable deadline. The broker retains each ticket in-process and advances only after actual native consumption. Only its exact consumed, unreleased prefix can explain held input at the next edge. The renderer must hold exactly those keys and no mouse button. Only then is physicalUp normalized for the original complete native scene comparison. Owner, actor, module, queue, window generation, focus, modal, capture and delivery-context checks remain active. Release/expiry/cancellation invalidates prefix authority. The root nonce is tombstoned even after partial failure; transport cancellation stops the remaining sequence. No partial replay or replacement authority is acquired.

The production controller now sends a native directional jump as one chord, retains every edge receipt, requires the complete reply inside its existing 500 ms bound and schedules independent UP from the first consumed edge. The fixed release cap, 1750 ms release check, 3000 ms operation, five-second game freshness and sticky terminal rules remain unchanged.

Focused native cases passed: compound success, old-helper isolated refusal, focus loss, actual guest MessageBox, actual public USER mouse capture, process identity change, queue replacement, foreign physical key, early release and delayed reply. Genuine modal/capture cases allow the authored guest to execute its public API before attempting the second edge. They prove first consumption followed by second-edge refusal, with independent UP and individual actual native/watchdog/outer/tool waits. Native single-key success, focus-loss release and delayed-reply cases also passed. The actual CLI guard test (including compound release, cancellation and owner changes), 19 connected send tests, receipt unit test, media manifest check and locked build passed. The build reproduced module 0d383140f6d4207d18860e3cb22fc8fcc47a5d260b8d5f2443222396b245fdd5.

Qualification failures are retained: an initial modeless color-dialog fixture did not create a modal; the next guest-API fixture attempted the second edge before SDL had entered its modal. Corrected fixtures explicitly wait for the authored guest's real API effect. The older controller-settle suite still intercepts queued png: commands although native capture now uses native-scene-capture; its four failed assertions/fault injections are retained, not counted as passes. An invocation without its required holder failed before launch. No production context check was weakened to accommodate these fixtures.

The affected original uses the coherent prior exact-image receipt graph with only lib/cli-guarded-input.js and test/run.js changed in its sealed runtime source manifest. All 789 media/settings pins matched. Runtime growth stayed in owned tmpfs with an independent finite watchdog and disk-floor cancellation.

The affected original visibly reached the title, bound the whole issued image, and consumed a personally selected Start Game Enter. A later personally inspected whole frame shows Slot 1 Free selected. Startup/draw-phase black captures authorized no input. A run-local duplicate review-record write failed before DOWN; a new complete capture/review preceded the actual Start Game input.

The Slot 1 Enter body was sent 9 ms after dispatch, but no response headers or reply arrived before the 500 ms timeout (error returned at 517 ms). Native consumption/effects of this Enter remain UNKNOWN. Independent UP acknowledged at 1762 ms after dispatch, exceeding the 1750 ms check by 12 ms; its actual release-child wait was 0/both EOFs. The scope permanently closed, with no later endpoint request. A subsequent capture invocation stopped at the local sticky closure check before sending. No unchanged retry or partial gesture replay ran. The original therefore did not reach gameplay or exercise the compound jump. No landing, course progress/completion, life-loss observation or ordinary Quit is claimed for this scope.

Closure was host cancellation: the native exit recorder retained .cancelled/SIGTERM, not an ordinary native Quit. Native, watchdog, outer-child, outer, driver-child and holder actual waits all returned 0/both EOFs; the finite supervision receipt reports no leak. All recorded runtime and supervision PIDs were absent. The detached driver's self-completion is retained, but its invoker did not record a parent wait; no such actual wait is invented. The messageboard watcher ended 143 separately from the native runtime.

The 119 original disk samples stayed above the floor (minimum 1,524,203,520 bytes). Shared disk capacity crossed below 1.5 GB during finalization after runtime closure; task-wide continuous compliance is false. Only this worker's closed evidence was compacted: 52 larger JSON/patch files have lossless gzip recipes, and 854 small qualification JSON records have a verified JSONL gzip archive with per-record SHA/length and original paths. Recovery instructions were durable before unlink. Selected original PNGs and unique authored inputs were retrieved and verified before tmpfs removal. Shared peer recovery restored capacity; no foreign evidence, media, history, DB or security hold was changed.

The remaining original barrier is the recorded no-reply confirmation timeout; without contacting the closed scope, the internal ticket outcome cannot be distinguished further. The proved two-key contract fix is published and native tested, while the requested manual landing/progress/Quit remains unmet. No boxes, workers, Wine, Python or deployment were used. Final box ID null, state NOT_CREATED, expiry null, stop receipt null.