Browser native input receipt bridge, 2026-11-17
The browser bridge is implemented and exercised through actual owning native USER returns on both current GPS backends. Original GLTron turns, round consequence and Quit remain unproved: the one original per backend closed at startup admission, with zero DOWNs. The measured admission/encoding faults are repaired and tested below. Both original allowances are spent; there was no original retry.
The opt-in browser bridge now carries a real renderer keyboard edge through the owning GuestWorker and confirms its exact public MSG at the native USER return. A browser/CDP ACK does not count. host.js installs the bridge into the actual import scope; lib/guest-rpc.js samples USER thread/process/focus/modal/window generations on the executing Worker, translates the MSG pointer there, and scopes the page broker to that slot. The bridge pins the process, instance, memory, transport, window, GPS endpoint/epoch/device/layer and source canvas. No callback runs on the page's idle instance.
Evidence is scratch/runs/20261117-gltron-browser-native-receipt-bridge-gameplay/. Exact Git and binary patches precede execution; module bytes precede the runtime clock and match both observed native/render Workers. The retained module is cdcbd50d245a9b3f6ff8c8ad9b61b52e24c9cf286c1f80dd045016cdbe42dd49. The candidate serves the immutable 97670c226df72bace657fcfefceba7938385d585 runtime plus only the four owned browser runtime changes. Other shared-tree WIP is excluded. Earlier lifecycle/module proofs are reused, with their exact source graph checked, rather than regenerated for unrelated HEAD changes.
The first complete authored runs candidate-software-v9 and candidate-webgl-v9 returned ArrowLeft through owning slot 0, thread 1, HWND 65537, PeekMessage with removal. Native receipts took 254 ms and 175 ms. Each independent process was armed before DOWN and observed a trusted physical UP with key state clear while the parent was stalled for 2200 ms. The 500/1750/3000 ms bounds are unchanged. Personally reviewed complete post-action host images show the red consequence; software includes the yellow triangle and WebGL's later host image is red without the triangle. Full source/output/host PNGs are retained, including different ordinary publications between ticket readback and host screenshot. These are not claimed to be an atomic three-image snapshot.
The production guard tests cover exact native MSG/token matching, wrong actor and import scope, focus, modal, generation, target/context replacement, closed/faulted endpoint, wall/monotonic deadline, untrusted DOWN, UNKNOWN, duplicate nonce and an enqueued edge followed by a throwing handler. A second matrix exercises the actual Worker encoder and page broker with authored USER export values and native pointer translation. These negative tests inject protocol context; they are not claims that original GLTron opened a modal or changed focus. Same-GPS positive runs exercise actual native USER exports and physical browser events.
WebGL's final initial and post-action complete captures finished within the original 3000 ms deadline. Bounded participant logs show the ticket retaining a held writer and then ordinary publication servicing it. The earlier published WebGL read-ticket timeout has not reproduced; no forced flush, repaint, larger deadline or blank replacement was introduced. Its old endpoint remains permanently closed.
An earlier authored native return trapped because the first bridge implementation tried to translate the guest MSG through a nonexistent page-context helper. Translation now occurs on the owning Worker and travels as the sixteenth broker argument; both actual native positives and the production transport regression pass. Earlier runs also preserve the refused expired ticket caused by arming a child after preparation; the release participant now arms before the 500 ms ticket exists.
Validation: focused native-input and complete-capture tests, Worker error-observer and caller-ESP tests, and the canonical locked build pass. The old import-throw test's obsolete direct-GL opcode expectation also fails against the unchanged baseline; its failure and baseline comparison are retained. The first two asynchronous build tool sessions returned SIGTERM despite complete logs; build-owned-exit.json records the replacement persistent owner's actual build wait with code 0.
The one registered software original on published 603cd76ff13955b873328781ec75fbcaf9c9568d exposed a startup admission fault. Its first initialized GL publication was a black loading display, before SDL's first native USER poll; the harness immediately called bind and correctly received a refusal because no actor scene had been sampled. It permanently closed with zero DOWNs and no later endpoint queries. Complete source/output/host images, module inheritance, all recursive media hashes and actual lifecycle waits are retained in original-software. This allowance is spent and is not retried.
The startup fault is repaired by passive readiness observation before capture/bind. The production bridge reports readiness only for a fresh owning native scene and a live matching window/GPS producer. The harness waits within the existing runtime budget, makes no bind attempt while unready, and closes permanently on timeout, transport error or retirement. No input, timer adjustment or idle-instance export establishes readiness. delayed-software-v10 and delayed-webgl-v11 faithfully publish native GL then execute a real 1800 ms guest Sleep before their first USER poll: both record unready observations followed by the actual scene. Software then returned ArrowLeft in 20 ms; WebGL returned it in 37 ms. Software’s independent trusted physical UP during the 2200 ms parent stall passed. The intermediate WebGL v11 UP observation took 1761 ms and is unqualified against 1750 ms; its parent-stall UP proof is taken from the materially matching v9 route (1712 ms). WebGL additionally returned a personally reviewed Escape and exited natively with code 0, without host stop; this is an authored Quit, not an original GLTron Quit.
The first delayed WebGL capture (delayed-webgl-v10) localized another wrapper fault: ticket complete at 1110 ms, host complete at 2649 ms, then synchronous PNG encoding pushed completion to 3052 ms. It closed with zero input. Those locally retained images are explicitly unqualified; no late endpoint query supplied a replacement. Evidence PNG encoding now uses one lossless row filter and deflate level 1. Full-size decode comparisons preserve every RGBA byte and alpha; no browser pixels or rendering are changed. The replacement WebGL initial capture completed in 1681 ms, and its action captures also meet 3000 ms. This fixes the measured encoding cost, not the older unreproduced published read-ticket stall.
The one WebGL original used published 806484714847faa48380bfecc5a3587bbdcdd96f, the same retained module, unchanged recursive media, registered defaults and real ticks under 300/320-second supervision. It observed native readiness, completed its whole black loading capture, then bind refused. The ordering left a capture between readiness and the bridge's fresh-scene check. Context age is the supported cause hypothesis; the precise final failing readiness subcondition was not sampled because the endpoint closed immediately, and no diagnostic query followed UNKNOWN. Its native/render Worker module inheritance and complete source/output/host images are preserved in original-webgl. Both originals have zero later CDP requests after failure and zero DOWNs.
The final repair makes observation and binding one synchronous page task before capture, with bind rechecking all fences. It does not lengthen the 250 ms scene fence or any 500/1750/3000 ms action bound. atomic-software-v12 and atomic-webgl-v12 faithfully add a second real 1800 ms native Sleep immediately after the first USER poll. Binding completes 1 ms and 19 ms after that native observation; subsequent initial captures take 504 ms and 367 ms. The old ordering is reproduced against the production guard by advancing the capture interval past 250 ms and observing refusal; the atomic ordering binds while fresh and never rebinds. Both actual GPS runs then return the exact ArrowLeft MSG (146 ms software, 91 ms WebGL), observe independent trusted physical UP, retain the personally reviewed red triangle consequence, and return a fresh Escape before native ExitProcess code 0. These are authored native Quits. Native receipt/UP-under-parent-stall proofs from the materially matching v9/v10 routes are reused. Both final initial source and output PNGs also pass the independent literal-coordinate oracle with zero mismatches over 172800 source and 788640 output pixels.
Final validation includes the production guard/Worker-broker negative matrix, passive startup/atomic bind regression, full RGBA codec/deadline tests, and canonical flock /tmp/wine-build.lock bash tools/build.sh with actual persistent-owned exit 0 (build-v12-exit.json). The original software first-GL-before-USER order is demonstrated and repaired; the original WebGL capture-between-observation-and-bind order is removed and faithfully tested. The repaired atomic admission has not been retried against an original. The precise remaining boundary is original menu/gameplay after admission: two visible turns, a round consequence and original native Quit all need a separately authorized original run. No claim of a guest GL crash, full game compatibility, FPS, audio, other apps or physical GPU follows from this work.
All native/browser/outer-owner/watchdog actual waits and both EOFs are retained. Original refusal cleanup is host teardown, while the two final authored runs exited through native Quit. No boxes, other Codex workers, deployments, source archives, media copies or frame bundles were created. The two foreign callback-review records remain exactly preserved as uncommitted shared-tree WIP; only the host review was published.
The final diagnostic-only correction records actual DOWN attempts and marks native sessions writable in the harness annotations. Earlier inputEvents:0 and readOnly:true fields were inherited labels; wire dispatches and native receipts remain authoritative. Focused tests were rerun and the final canonical locked build has an actual persistent-owned exit 0 (build-final-owned-exit.json). These annotation changes do not alter the qualified native or capture paths and did not consume another original allowance.