SuperTux reviewed native capture binding, 2026-11-18

Evidence: scratch/runs/20261118-supertux-reviewed-native-capture-binding-course/.

Commits e22a60149 and f86743427 repair the native-scene capture/bind protocol. The old controller took its native snapshot and PNG in separate requests; bind then encoded a later repaint and compared that new image's hash with the reviewed image. The retained predecessor explicitly refused reviewed native paint changed before creating an owner or DOWN. Its scope remains closed and was never contacted in this task.

The new server action captures the full display and native context in one synchronous dispatch. It retains the exact encoder bytes and issues an immutable random capture identity, SHA-256, original times and a 30-second expiry. Both wall and monotonic clocks enforce expiry. Bind verifies the issued record and exact image association, the retained hash, current native process/window generations, focus, modal, capture and physical state, all native window metadata, and the original renderer/queue/actor/module/device references. A changed reference cannot acquire a replacement authority through matching public metadata. The existing permanent owner and consume-time native checks remain in force. The legacy no-receipt route retains its hash comparison. No pixel crop, mask, repaint freeze, timer change, review renewal or input-deadline extension is used.

The controller verifies the saved whole PNG against its receipt and the personal review hash before binding. Its existing 5-second game freshness, 30-second menu freshness, 500 ms confirmation, independent physical release, 3-second operation bound, sticky UNKNOWN and no-replay rules are unchanged. LMarbles released controller ownership before these edits; common and supervisor files were read-only. The 786dae2f8 SuperTux manifest regression continues to pass.

The first original returned a black whole PNG, while a later diagnostic queued whole capture showed the title. It closed with zero owner/DOWN and all recorded waits 0/both EOFs; cancellation is not ordinary Quit. Commit f86743427 made the compositor call explicit inside capture, following the queued PNG path. Final inspection found that the earlier action already called presentForRead(), whose hook also composites. Therefore an omitted composite is not a proved cause of that black image, and the lazy-compositor test does not distinguish the two versions. This commit is explicit compositor alignment, not an independently proved second defect. The second original also produced black startup/draw- phase captures before visible frames; those images never authorized input. The proved protocol defect is the separate snapshot/image and later-repaint SHA comparison repaired by the server-issued exact-image receipt.

Actual native SDL/HTTP tests use an independently authored animated PE, real SDL fills/flips, the production capture/bind/controller path and identical input context across changing whole images. They cover genuine different capture substitution, real 30-second staleness, process owner, focus, modal, window replacement, queue replacement, real public USER mouse capture, independent release after focus loss, and a genuine delayed consumed reply that remains refused/UNKNOWN under the unchanged 500 ms bound. The final composite regression suppresses incidental display composites and requires the capture action to invoke the real compositor and return a nonblack whole PNG. Ten cases passed in that suite. Its UNKNOWN fixture suffered a first- capture ECONNRESET before input and a native SIGTERM teardown; those failures are retained. A fresh finite UNKNOWN fixture passed separately with independent UP and all actual waits 0/both EOFs. Earlier fixture assertion/interception mistakes and their actual waits are preserved as failures, not hidden passes. The receipt unit tests, 19 controller tests, manifest regression and required locked builds passed. Each code commit was immediately strictly pulled/pushed. The final build reproduced module 0d383140f6d4207d18860e3cb22fc8fcc47a5d260b8d5f2443222396b245fdd5.

All 789 original media/settings pins matched before launch. No media, settings, save, game state, clocks or private memory were used to choose gameplay. The 22 predecessor recovery archives and recipes and all holds remain intact. Only this worker's closed test metadata was compressed, with disk roundtrip, SHA/length/stability and recovery note before unlink; individual recipes are retained. Actual disk observations briefly crossed the 1.5 GB floor during qualification/preparation; no task-wide continuous compliance is claimed. The first original's sampled minimum was 1,504,604,160 bytes. LMarbles then reported a separate verified closed-log recovery restoring capacity.

The materially repaired second original is recorded separately under original-composite/, with an unchanged 1200-second runtime and 180-second Quit reserve, independent finite supervision, dedicated native exit receipt, sealed source/module inputs and exact operational helper hashes. Its gameplay and exit outcome follows. A genuine whole title capture bound successfully while normal animation continued. Menu actions selected a new game, the introduction, and the world-map node “Welcome to Antarctica”. Personally chosen finite Right and Space inputs advanced along the ice and visibly earned one coin/25 points. No life loss was observed. Several intentions expired before DOWN and remained NOT_SENT. World-map course selection used the menu freshness class; actual course movement/jumps used the unchanged five-second game limit.

A final run-local two-key intention was invalid: Right was consumed, then Space was CANCELLED with “reviewed native scene changed”, because the first held key changed the reviewed physical context. This is a real refusal, not harmless repainting or a receipt failure. The run-local helper marked sticky terminal failure and requested cancellation; no new endpoint request followed the already-prearmed release. Its prearmed independent release acknowledged UP and exited 0 with both EOFs. The definite CANCELLED reply is not an actual UNKNOWN, but the conservative terminal closure was honored. No unchanged retry followed. No platform landing, course completion, or ordinary Quit is claimed.

Native, watchdog, supervisor-child and outer actual waits were all 0 with both EOFs. The independent native-supervise receipt reports exit 0/no leak. The native exit receipt is explicitly .cancelled/SIGTERM with the invocation token, not a native Quit. The enclosing run-local driver tool instead ended 143; its final wait/disk/holder records were not written, and no successful enclosing wait or continuous second-run disk trace is invented. The recorded native, holder and enclosing PIDs are absent. The recovery recipe retains the original module archive association; no temporary memfd is required for recovery.

The protocol repair is published and demonstrated on animated native input. The requested full platform/course progress and ordinary Quit remain unmet. There were no boxes, workers, Wine runs, Python calls or public deployments. Final box ID null, state NOT_CREATED, expiry null, stop receipt null.