GLTron atomic control and native round, 2026-11-13

The parent request-publication race is fixed in a new run-local helper. The original local game visibly made one yellow-cycle corner and finished a native round with WINNER: 3 / PLAYER 3 WINS!. Two distinct controlled corners remain unproved: both S attempts were acknowledged, but their reviewed frames show Player 1 already crashing or eliminated. This is a concrete gameplay-observation failure, not evidence of a production renderer or interpreter crash. No production files or parent artifacts were changed.

Evidence: scratch/runs/20261113-gltron-atomic-0230/ (ignored). The tested main SHA is 797f19cdb4a5bfb1650fffd6b516311bd2c1205d. Preflight and postflight verified all 1,617 parent source paths and all 92 existing media paths by size/hash identity, with zero mismatches. The existing 2,189,863-byte module remains SHA-256 04c534192ee3fe39e388536c2ce710c12c47a5cd2ca4d95c19f9d00b1e8f0a17. The parent source-pins.json and media-pins.json are retained by path and SHA in this run's parent-references.json; media was neither copied nor staged. Tracked relevant runtime diff is empty; the existing untracked runtime addition is preserved as untracked-runtime.patch. No source archive exists.

Run-local repair and focused tests

control.py publishes each request to an exclusive random temporary file, flushes and fsyncs it, then renames it into the ready namespace and fsyncs the directory. A permanent exclusive identity reservation refuses publisher reuse. The consumer ignores temporary files and renames each ready request to a permanent claim before parsing or sending. Duplicate claimed IDs cannot send again; malformed JSON and mismatched/invalid commands receive explicit invalid receipts without taking down the consumer. Commands preserve their original identity, and actual runtime replies create separate ACKs. Pending or lost requests are never automatically resent.

Each finite key pair arms its UP timer before DOWN under the send lock. The UP timer runs independently of request polling and ACK handling; cleanup also attempts release of held keys. Every actual key pair in this run received both ACKs (14 key ACKs total). A timer duration is a host send interval, not a claim about guest-observed hold duration.

tests-complete.log records five passing tests: partial publication ignored until rename; duplicate publication/consumption refused; malformed and invalid identity handling; ACK versus terminal UNKNOWN/noReplay; an actual exited child with a broken stdin pipe; and signal cleanup with a real substituted child, actual wait and both separate EOFs. The earlier four-test run had a test teardown BrokenPipeError on closing the already-broken pipe; it is preserved in tests-final.log, and the corrected teardown passes.

executed-control.py preserves the exact helper used for the game. The final control.py additionally distinguishes SENT_UNCONFIRMED from terminal UNKNOWN and handles TERM/INT through cleanup. Those final changes passed the five-test suite, including the real signal/EOF test; they were not rerun against a second game runtime. All helper versions/tests are preserved as addition patches. No production code changed, so no production build or broad matrix was run. The parent native16/GL/D3D investigation and already-passing repairs were reused; no new production failure justified a runtime patch.

Original local runtime and reviewed frames

The fresh route was the parent's ordinary test/run.js --app=gltron --no-build --no-threads --gl-renderer=software --real-ticks --batch-size=200000 --max-batches=1000000000 --quiet-api --quiet-blocks --gl-census --control-stdin --png-canvas route, with internal max-seconds 240. GNU timeout was prearmed before launch with a 260-second TERM deadline and five-second KILL fallback. deadline-armed.json and launch.json preserve exact command, time and wrapper PID872444. Runtime PID872445 was checked by its exact command before stopping. Initial actual /home free space was 1,504,477,184 bytes, above the required 1,500,000,000 floor. No media staging, RAM gate or housekeeping occurred. No API, browser or box was used.

Full 640×480 frames were individually inspected before each next finite input:

The second-turn observation failed twice during live frame-review latency. No tracking, automatic path or blind action loop was used. No guest memory, configuration, AI or clock writes occurred; the original real guest clock ran throughout. Native cycle collisions above are ordinary game events.

Stop and limits

The tool-owned supervisor session ended143 before the final finish request was claimed. The cause of this signal is unknown; it is separate from the fixed JSON publication race. Its exit left the independent timeout and runtime alive. Both were identified, and the exact owned runtime PID872445 was sent TERM at 00:25:41Z, before the independent deadline. manual-stop.json records the command and signal; postflight.json verifies both runtime and wrapper absent. All key pairs had already received UP ACKs. No inherited request was replayed, and unconsumed finish is explicitly UNKNOWN/noReplay.

Actual runtime wait status and separate stdout/stderr EOF are UNKNOWN; no terminal receipt or final PNG is claimed. The later synthetic-child signal test proves final helper cleanup behavior, not this game's exit status.

Box ID none; state NOT_CREATED; expiry null; box stop receipt N/A. No inherited box was changed. The task stops with the race fix tested, one observed corner and native round outcome, and a distinct failure to prove the requested second corner. Next work, if assigned, is a fresh bounded ordinary-input observation that captures a second native turn before collision; these S ACKs must not be promoted into visual proof.