Nox complete transport preflight and native delivery blocker, 2026-11-07

No new original Nox gameplay or equipment evidence was obtained. The corrected transport passed positive Chromium request/consume/native-key tests locally. The sole fresh remote driver then terminated during its native 50 ms keyboard preflight, before original selection or launch, with AssertionError: input delivery exceeded hold bound. This is a harness delivery deadline failure, not evidence of a Nox gameplay defect. The first blocker ended the attempt; there was no replay, replacement box or second remote driver.

Evidence: scratch/runs/20261107-nox-complete-transport/.

Corrected transport and actual local preflight

The staged transport.cjs now imports the actual staged ordinary-keys-fixed.cjs, fixing the previous missing ordinary-keys.cjs dependency. Its request path also checks the returned ACK ID against the requested ID; the underlying generic ACK client did not enforce that identity. Both changes are run-specific harness inputs retained as patches, not changes to the emulator or shared production helpers.

generated-corrected/positive-transport.json records successful numbered key and combo requests through the exact staged transport, atomic publisher, consumer, ordinary adapter and native Chromium input. The key is I; the combo is Up+Left. Both produce the exact expected ACK and trusted DOM down/up events, empty physical sets and cleared renderer keyboard bitmap entries. Releases complete while a 300 ms diagnostic is still pending for an 80 ms hold. These are generated-fixture proofs of transport operation, not original gameplay. The labeled generated canvas was inspected.

The fixture retains five process/endpoint/Worker/module/epoch replacement refusals, stale/slow capture refusal, native partial delivery cleanup and independent release. UNKNOWN and late-ACK tests now require the exact ACK deadline; never resend error, actual published command and sticky closure artifact; subsequent requests require the intended closure error. Stale refusal requires the specific fresh-capture-only error, partial delivery requires its injected native-delivery error, and a wrong ACK ID explicitly closes the transport. Missing modules and unrelated exceptions fail these tests. The initial local fixture correctly failed because its expected legacy-key error text was inaccurate; the predicate was corrected to the actual assertion and the complete fixture reran successfully. That failed fixture is retained.

The exact browser-run-fixed.cjs production entrypoint also ran locally in Chrome 151.0.7922.108, including native key/name/combo/partial preflights, HTTP pin checks and final Nox app selection. Only its fixture environment flag suppressed launch; runningApps.length was zero. The flag was absent remotely. The declared 765-file pin/deadline prefix passed all missing/corrupt-file and invalid-deadline controls. The actual staged filenames and positive transport execution replace the predecessor's false-positive broad rejection checks. Three focused ops/browser-native-input.test.cjs tests passed. The locked build passed after the attempt. No shared emulator or helper code changed.

Sole fresh attempt and exact blocker

UUID-idempotent box bx_waf55qq6 was created at 09:53:00 UTC. The inclusive 35-minute deadline was 10:27:59.858 UTC, browser cap 10:21:59.858 UTC, initial archive expiry 10:33:00.058 UTC. All 765 uploaded pins matched before execution. One driver and Chromium started at 09:53:23 UTC. At 09:53:24.113 UTC, the native helper rejected delivery in the driver's first key-preflight loop, line 99: bounded detected that its independent release had already started before the keyDown CDP call returned. The guard remains intact; it was not weakened. The log does not identify which loop key failed, and no timing receipt was persisted for that failed call. Exact remote DOM delivery counts are unknown.

There is no launch.json, selector record, guest process/Worker/module identity, original renderer/backend, game input or gameplay command. The generic launch-parent.json field originalLaunches:1 counts the driver attempt; actual original guest launches are zero, as terminal inspection establishes. No Multiplay/Host/Bunker, equipment, pickup/HUD/action consequence, target damage, FPS, audio or parity claim is made. Balanced inputs alone would not prove those.

Pins, recovery and lifecycle

All 125 original files were rehashed locally against the retained original manifest: 50,450,550 bytes, exact paths with no extras/missing files. EXE SHA-256: 496123ad275e5d626f31a845033c7ddb7d86fbbf3c2532d2e1f639f8bc77a781. The tested source commit is af80bc10f507c9722d8b5d5c1fcc4fa29151db29, also recorded in source-commit.txt; relevant dirty source hunks and owned run inputs are retained as binary Git patches. Git is the source archive; no source/media archive was made.

The standard module is 2,182,493 bytes, SHA-256 e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467. It matched local and remote HTTP/pin checks and was recovered independently as tested-module.wasm. It was never a live original guest module in this run. Remote Chrome was 151.0.7922.108, V8 15.1.206.16, native inner geometry 1024×712, defaultViewport:null, without viewport emulation.

All 31 remote artifacts, 2,394,714 bytes including the module, were retrieved and SHA-256 verified before delegated normal nonforced stop. Terminal checks verified both controller/Chrome PIDs absent, both HTTP/DevTools ports refused, no Chrome processes, profile or writer lock. Original input recorder and command directory were empty; input clearance is based on native helper finally/timer cleanup and complete browser termination, not a remote DOM balance receipt. Chrome exited code 0. The box is archived, stopped at 09:53:58.822 UTC, expiry null, stop completed with forced:false. Ownership returns to the coordinator; this terminal session must not be resumed.

Local free disk exceeded 2 GB before allocation; shared activity later reduced it to 1,833,259,008 bytes after all runtime had ended. This worker removed no peer files and launched no further runtime. The incident is retained; the final local check recovered above 2 GB free (2,629,058,560 bytes).

Next work is a separately authorized repair/qualification of cold-browser native delivery versus the 50 ms preflight interval, preserving independent release and exact admission guards, followed by fresh original equipment/action exploration only after that passes. The corrected positive transport evidence is reusable input coverage; it does not resolve the new remote timing blocker.