Original GLTron: actual native GL live turns, 2026-11-14

The registered original GLTron rendered through actual native --headless-gl, made two reviewed live A turns in the final fresh round, and reached the natural PLAYER 3 WINS! result. No production source was changed. Evidence is in scratch/runs/20261114-gltron-native-gl-turns/result.json.

Actual renderer and inputs

The passive observer records one successful native context, successful binding of context 1 to owner 0 and hwnd 65537, currentByOwner=[[0,1]], and gl_sw_enabled=0. Authentic guest presentation commands produced native readbacks, initially black during asset loading and then nonblack. The run completed 2,952 authentic presents/native readbacks, 460,592 executed draws and 18,189,063 vertices. No later context creation or binding switch appears in the trace. The selected compositor images were individually viewed. This is actual native WebGL on llvmpipe LLVM 20.1.2; it is not physical-GPU, browser, parity or historical 27-app acceptance.

The original 92 media paths were unchanged. The actual module read was 2,199,488 bytes, SHA-256 c7f5bada4da4b4ed95b235add0aa1c9843176805d29d5b1ee79359752a19d348. The tested main commit was 3ea50a0f34264ebd94d9c7a908e54d3eb5a2571f. All 317 launch source pins remained unchanged; all 300 overlapping predecessor pins matched. Git plus the retained binary source diff and addition patches preserve the dirty runtime identity, without a source archive or module dump.

The command used the registered app, --no-build --no-threads --headless-gl --real-ticks --max-batches=1000000000 --max-seconds=240 --quiet-api --quiet-blocks --gl-census --control-stdin --png-canvas. No batch-size override was supplied: the default was 1,000 blocks. The observer wraps existing context/bind/present/readback/compositor functions without additional rendering, readback, swap or forced-presentation calls. No software fallback was selected.

Reviewed final round

All image names below are relative to the evidence directory.

Each selected turn followed review of the current intact cycle, forward path and minimap; focus remained hwnd 0x10001. The control publisher reserves IDs exclusively, claims before sending, never replays UNKNOWN, and prearms each independent finite 80-ms UP timer before DOWN. ACKs establish release and batch progress; the viewed geometry establishes turns. There was no route automation, tracking, guest-state/configuration/clock/AI write or synthetic GL publication.

Earlier rounds are explicitly rejected as turn proof. The initial start/pause commands were consumed during transition and the first round reached PLAYER 4 WINS. In the next round, attempting Space as resume produced no reviewed corner; a later settled frame showed yellow crash debris (native-corner-one-settled.png). That round was allowed to finish naturally before the final fresh round. Return was then used to resume from the visually verified native Space pause, as in the successful predecessor. No failed or dead turn is counted above.

Recovery, checks and lifecycle

The predecessor had removed its regenerable native install. Its retained npm lock and offline cache restored glfw 7.4.1/webgl 6.0.3; all 162 dependency file hashes exactly matched. Initial native preflight failed the resized-buffer right-edge readback assertion. An unchanged repeat passed the complete native shader/readback/compositor test. Both logs are retained: the intermittent resize failure remains a limitation, not a claimed production repair or hidden PASS. The actual original subsequently rendered through the observed native backend.

Eight focused input/lifecycle tests passed, including balanced Space, duplicate and invalid refusal, actual lost-child UNKNOWN/no replay, normal completion, supervisor signal and independent deadline expiry with actual waits and both EOFs. These synthetic tests are controls, not gameplay evidence.

One immutable actual gameplay attempt ran to the original 240-second bound. Its independent 260-second watchdog armed before the runtime child. Runtime wait was 0, stdout and stderr both reached EOF, watchdog wait was 0, both PIDs were absent afterward, and every key DOWN had its matching acknowledged UP. The heavy runtime lock was released at completion; no lock is held during finalization. Builds additionally use the build lock. The first finite build queue expired without launching a build. The next fair queued turn actually ran tools/build.sh under both locks and failed at the foreign-owned ToyVM browser bundle freshness gate: 2,651,545 bytes versus 2,651,735 regenerated bytes, first differences 801005/853049. build.log and build-receipt.json retain that failure. No passing full build is claimed, no owner files were edited, and this publication contains documentation only.

Only seven selected result images, compact logs/JSON and small Git patches are retained. The separately viewed redundant captures were removed. The restored native dependency install was hash-verified before its own regenerable files were removed; the lock/cache recipe and pins remain. No media was copied or downloaded. Free /home disk remained above 14 GB, exceeding the 1.5 GB floor. No box was created: ID null, state NOT_CREATED, expiry null, stop receipt N/A. Foreign WIP and files remained untouched. The assigned gameplay task is complete; no further runtime attempt is needed.