Nox exact-control repair and prelaunch blocker, 2026-11-07

The missing I/Home/Space support is repaired and qualified in actual Chromium. No new original gameplay or equipment evidence was obtained. The sole fresh driver terminated during a negative-duration preflight, before Nox was selected or launched. This is another harness defect, not an emulator or game defect. The previous terminal run 20261107-nox-weapon-interaction and its command 31 UNKNOWN result remain untouched; neither session was replayed or resumed.

Evidence is in scratch/runs/20261107-nox-qualified-equipment/.

Input repair and verification

Published helper ops/browser-native-input.cjs centralizes checked CDP key descriptors, required-control assertions, bounded key combinations, and left/right mouse holds. I is key:i, code:KeyI, VK 73; Home is VK 36; Space is key:" ", code:Space, VK 32. Key/button release runs on an independent timer and in finally, including partial delivery. A slow diagnostic or persistence callback cannot defer release. Physical host intervals do not prove guest polls or steps.

The new driver generator checks every replacement anchor and replaces the bound helper's complete hold implementation with the shared helper, rather than silently trying to extend a textual key-table suffix. The outer ordinary wrapper uses the same descriptors and delegates to the actual bound helper. Required I/Home/Space and navigation controls, plus both bound input methods, are asserted before any original launch. Arrow-only qualification is no longer sufficient.

generated/exact-controls.json reproduces the old mismatch in Chromium: the legacy ordinary wrapper validates each of I/Home/Space, then the actual legacy bound hold rejects each before dispatch, with zero native events. Its after rows independently exercise all five required controls through the bound API and through the ordinary wrapper delegating to that bound API. All ten paths produce exactly one trusted down/up pair. I/Home/Space have the exact DOM key/code values; released keyboard bitmap entries are zero. Both mouse buttons produce native button pairs. Each path uses a 300 ms diagnostic with an 80 ms hold; native release completes while that diagnostic is still pending, and all completed samples are explicitly excluded from in-hold progress evidence.

Both generated/ and generated-corrected/ pass the actual browser chain's five owner/process/endpoint/worker/module/epoch replacement cases, stale and slow-capture refusal, independent release, native partial delivery, sticky UNKNOWN and late ACK closure. Those tests use generated fixtures only; their worker/module identities are not original Nox identities. The generated labeled canvas capture was inspected. The three focused reusable-helper tests and flock /tmp/wine-build.lock bash tools/build.sh passed before helper publication d10ebc5d6 on main.

Exact first blocker in the fresh driver

Box bx_u7qt7nrc was allocated with a persisted UUID idempotency key at 09:36:53 UTC. The authorized 35-minute window ended at 10:11:53 UTC; the box's initial expiry was 10:16:53.940 UTC. There was one driver/Chrome attempt.

At 09:37:29.345 UTC, browser-run.cjs line 101 tested invalid hold durations using assert.rejects. The new outer ordinary-keys.cjs wrapper performed validation synchronously, so an invalid-duration assertion escaped before a promise was returned. The inherited wrapper had an async API. That compatibility mistake was introduced in this task; exact valid-key tests did not catch it. original/first-browser-failure.json and runtime-browser.log retain the first fault. The original selector/navigation code was never reached. No original commands exist, and no original native-input recorder events exist. The generic launch-parent receipt labels the driver allocation originalLaunches:1; the separate launch-accounting.json corrects this to one browser attempt and zero original guest launches, without rewriting the receipt.

ordinary-keys-fixed.cjs restores an async holdBound API. generated-browser-fixed.cjs repeats the actual Chromium checks and additionally executes the exact inherited invalid-duration assertion loop: six rejected promises, zero native effects. generated-corrected/invalid-duration-contract.json and its controls pass. browser-run-fixed.cjs is a reviewable corrected next driver, not executed against an original game. No original attempt was restarted after the fault. A future original run requires fresh authorization.

Closure, evidence and lifecycle

All 125 originals match the retained SHA-256 manifest and shipped FILESIZE.INI paths/sizes: 50,450,550 bytes, preserving original layout. The registered 124 data files plus separately loaded nox.exe cover that closure. Original EXE SHA-256 is 496123ad275e5d626f31a845033c7ddb7d86fbbf3c2532d2e1f639f8bc77a781. All 764 staged runtime/input pins matched before the attempt and after teardown. The tested source commit is recorded in source-commit.txt; relevant dirty and owned new inputs are retained as patches. Git remains the source archive. No source/media archive was created.

The current standard module is 2,182,493 bytes, SHA-256 e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467. The preflight HTTP resolver fetched and verified it. There was no live original guest module, process, guestWorker, renderer owner/backend or scheduler to observe. Chrome retained native viewport geometry, defaultViewport:null, with no emulation. These facts do not establish any Nox rendering or gameplay result. No FPS/audio/parity, equipment, pickup/HUD/action or damage claim is made.

Chrome exited normally with code 0 and no signal. The controller exited after the first failure. Independent terminal checks confirmed both PIDs absent, both server/DevTools ports closed, no Chrome processes, no profile/writer lock, and no held recorded input. All 29 unique remote artifacts (206,356 bytes) were retrieved and SHA-256 verified before explicitly delegated normal nonforced stop. An initial local retrieval comparison caught a formatting-only newline difference in a locally reconstructed cleanup JSON; matching the remote bytes allowed the complete verification. No runtime or input was replayed for retrieval.

The box was then stopped normally; the final state and completed nonforced stop are recorded in local-cleanup.json. Local free disk stayed above 2 GB at the recorded checks. The stopped box is handed back to the coordinator and must not be resumed. Next work is a separately authorized original run using the corrected driver and its exact prelaunch assertion coverage, then ordinary equipment exploration. Target damage still requires an actually observed target.