GLTron two live turns and natural round result, 2026-11-14

The original registered GLTron reached two distinct reviewed live native turns in one fresh round, followed by the natural PLAYER 3 WINS! result. Three individually chosen finite A turns were sent in the successful round: the extra third turn provided another live heading/path view because the lower horizontal minimap segment was difficult to see. This is original gameplay, not fixture acceptance. No production source was changed.

Evidence: scratch/runs/20261114-gltron-two-live-turns/result.json, with the successful session in attempt2/result.json. The tested main SHA was a9448409727cc11d065591831103f03ae1b71c22; the 2,199,488-byte runtime module SHA-256 was c7f5bada4da4b4ed95b235add0aa1c9843176805d29d5b1ee79359752a19d348. The 92 existing media paths matched the parent hashes before and after; launch-source-pins.json and postflight.json establish that the relevant source files did not change during the successful session. Git plus the retained binary diffs and addition patches preserve the source identity. No source archive, media copy, module dump or capture bundle was made.

Reviewed successful round

All paths below are relative to the evidence directory's attempt2/.

Before each finite turn, a full frame's cycle, forward path and minimap were reviewed, and dump-focus reported the same main hwnd 0x10001. Each selected turn used a finite independent UP timer, followed by a chosen settling interval and ordinary Space pause. There was no automatic route, tracking, blind action loop, guest-memory/configuration/clock/AI write, browser, box, media provisioning or fixture promotion.

Rejected attempts and controller correction

The earlier 0300 note and fcdde27ce atomic0230 note were read. Their original one-corner proof and second-after-crash failure were not promoted into this result. The pruned helper and lifecycle tests were reconstructed from their surviving small Git addition patches.

The first fresh session incorrectly used Return as if it paused during play. The reviewed images lacked a pause banner. That session's second turn is rejected; its final image shows a natural PLAYER 3 WINS result but does not prove the requested live sequence. This was a controller/observation mistake, not a production defect. It ended with normal quit and complete wait/EOF.

Original executable disassembly, retained as native-space-pause.txt, proves keyboardGame at 0x408631 compares SDL key 0x20 (Space), branching at 0x408635 to0x4086a7, which sends RETURN_GAME_PAUSE=2. scripts/main.lua maps that return to the native pause callback. The run-local allowlist was extended only to Space, and the fresh session visually verified GAME IS PAUSED after Space. Its initial round still rejected S as a second corner because the settled frame showed yellow crash debris. That round was resumed without turning to its natural PLAYER 3 WINS outcome before starting the successful fresh round.

The recovered publisher retains exclusive single-use ID reservation, exclusive temporary publication, fsync/rename, permanent claim-before-send, duplicate/invalid refusal, and independent UP timers armed before DOWN. ACKs are separate from terminal UNKNOWN, which is never replayed. The only command extensions are read-only focus inspection and balanced native Space. first-executed-control.py preserves the first session's helper; control.py and addition patches preserve the corrected one.

Seven focused controls/lifecycle tests passed in focused-tests.log, including actual exited-child/no-replay, normal quit, supervisor signal and independently expired deadline with actual parent wait and both separate EOFs. After the Space allowlist correction, five focused controls passed in space-tests.log, including balanced Space and rejection of an unsupported key. Synthetic process tests are not gameplay evidence. The full build passed under both the heavy-runtime lock and flock /tmp/wine-build.lock bash tools/build.sh; build.log records it. It reproduced the exact tested module hash. Production source was not repaired or claimed.

Lifecycle and scope

Both native sessions held /tmp/wine-heavy-runtime.lock. Each separately observable watchdog durably armed before its direct runtime child launched, with an independent 260-second deadline and exact PID/start-token targeting. The runtime's own bound was 240 seconds. Both sessions ended normally through quit: actual runtime wait 0, stdout EOF and stderr EOF, watchdog wait 0, and runtime/watchdog PIDs absent. The successful session stopped at 09:41:08Z within its original bound; no deadline was extended. Every actual key pair received DOWN and UP ACKs before quit, and final-turn postrelease progress is recorded independently in postflight.json. Heavy lock was released.

/home remained over 14 GB free, above the 1.5 GB floor. Foreign WIP, TIM adapters and bundles, graphics source ownership and inherited artifacts were preserved. Box ID none; state NOT_CREATED; expiry null; stop receipt N/A. No new work is requested: the assigned single gameplay task is finished and stopped.