GLTron atomic original turns and round, 2026-11-17
Evidence: scratch/runs/20261117-gltron-atomic-browser-original-turns-round/.
This task reuses the published owning GuestWorker native USER receipt bridge and the corrected final software/WebGL proofs from 20261117-gltron-browser-native-receipt-bridge-gameplay. The intermediate WebGL v11 physical UP at 1761 ms remains unqualified. The 500 ms native receipt, 1750 ms independent physical UP, 3000 ms whole capture, 250 ms scene freshness and 15/30-second image fences are unchanged. This task's explicitly authorized native/watchdog runtime is 420/440 seconds; the CLI deadline admission now accepts that runtime.
Before any original, the complete actual CLI is exercised with the authored native GL executable through immutable preparation, persistent process owner, independent watchdogs, startup passive observation and atomic bind, whole-image review, Enter, ArrowLeft and Escape. These are authored key-route tests, not original menus or gameplay. Software records native receipts at 72/12/32 ms and physical UP at 143/75/101 ms; WebGL records native receipts at 89/111/17 ms and physical UP at 171/216/80 ms. Both return native ExitProcess code 0 without host stop. Selected complete host images were personally reviewed before every finite intent. All native/browser/outer/watchdog waits and both pipe EOFs are retained.
An earlier software authored rehearsal delivered Enter in 136 ms but cancelled the following ArrowLeft before any native token. Its generic owner/focus/modal/generation/target/deadline reason does not identify the exact predicate. The endpoint closed permanently; no query followed failure. Its independent trusted physical UP arrived 1700 ms after DOWN. This is retained as an unlocalized authored refusal, not an original failure or a claimed repaired runtime fault. The guard now retains the individual failed-fence observations at the first cancellation; no predicate or deadline changed. A focused regression crosses the existing 250 ms scene fence between preparation and native polling, checks the exact diagnostic, and proves a later native observation cannot revive or rewrite the cancelled edge. That regression demonstrates diagnostic behavior, not the cause of the earlier run.
The one software original used published ea6bbf6b8, unchanged recursive media and the exact retained cdcbd50d245a9b3f6ff8c8ad9b61b52e24c9cf286c1f80dd045016cdbe42dd49 module. Passive startup and atomic bind passed. The initial whole image shows black loading; the next complete image shows the original GLTRON title, yellow/blue bikes and GAME selected above VIDEO/AUDIO/QUIT. This is real original menu progress, not gameplay. Its personally reviewed Enter intent was refused during preparation before any DOWN: native scene age 589 ms, with healthy owner, valid scene, unchanged references, visible matching process/instance/window, matching scene/slot and both ticket deadlines unexpired. Software's allowance is spent; it is permanently closed and is not retried. No later parent CDP or HTTP request followed the refusal. The prearmed independent process sent its safety UP, but no DOWN existed to generate a qualified physical-UP receipt; its secondary code 1 masked the primary preparation error in the old wrapper. This is host cleanup, not native Quit.
The working-tree repair passively waits for a new actual native USER observation after the release participant arms, then prepares synchronously in that same page task. Startup binding is reused, never repeated. No expiring ticket or DOWN is created while stale. The existing 250 ms freshness, image expiry, runtime deadline and independently armed UP deadline all remain binding. The wait ends before the UP participant has less than 500 ms remaining. A faithful production-guard test reproduces the exact 589 ms refusal, verifies no ticket from the stale observation, then returns an exact native MSG after a new observation and atomic preparation. Refused nonce tombstones remain cancelled. Actual software CLI rehearsal (fresh-software) returns Enter/ArrowLeft/Escape in 20/24/19 ms, independent UP in 119/91/116 ms and native ExitProcess code 0. A separate actual release participant cancels before DOWN with zero input dispatches and actual wait/both EOF. The wrapper preserves a primary admission error when release cleanup also fails.
On WebGL, fresh-webgl returns Enter/ArrowLeft in 168/117 ms with independent UP in 302/207 ms through the repaired admission. Its next complete capture times out at 3002 ms while reading the ticket, so it is not a complete new CLI qualification or a native Quit proof. A separate atomic-intent-webgl authored instance times out on its initial ticket read at 3005 ms with zero DOWNs. Both remain closed with no later queries. No forced publication, render, guest pump or larger deadline was used; the old read-ticket timeout is now reproduced at the same phase, but a common cause or repair is not established. The earlier completed WebGL authored Quit is reused only for its actual older route, not relabelled as a completed new run.
Another separate authored control (final-webgl) read a command file during its non-atomic local creation and refused partial JSON before any input. The command recipe now publishes closed .pending files by atomic rename. This is an operator/procedure fault, not a native runtime fault or an excuse to retry an original. WebGL's one original allowance is unused because a complete current-backend rehearsal has not qualified. Original two live turns, round consequence and native Quit remain unproved on both backends. No source archives, copied game media, new sandboxes, other Codex workers or public deployment are used.
Publication boundary: ea6bbf6b8 publishes only the refusal diagnostics, the 420-second CLI admission and the earlier authored complete-route evidence, after its focused tests and canonical locked build passed. The subsequent fresh-scene preparation repair in tools/gltron-current-original.cjs, tools/gltron-current-native-input.cjs and test/test-gltron-current-native-input.cjs remains uncommitted and unpublished. Its focused tests pass, but the required canonical build is blocked by concurrent HL2 Edit WIP failing the control-variant attribution gate (edit_layout_matches, edit_layout_build_page, native_edit_viewport, and obsolete edit_layout_build attribution). Those foreign files were not edited or reverted. The failed build and timed gate wait are retained. The repair must pass the canonical locked build before a future worker may commit it; the WebGL capture stall additionally blocks a current complete qualification. No further original was launched to work around either boundary. All eight task-native scopes are closed with actual native/browser/outer-owner/watchdog waits and both EOFs verified; original closure is host cleanup, not native Quit. No box was created (ID/expiry null, state NOT_CREATED, stop receipt NOT_APPLICABLE).