Nox native readiness: reuse the durable release service

Evidence: scratch/runs/20261110-nox-native-readiness-cause-0119/. Prior fixture repair: visible pointer publication.

The retained failure is before CDP or guest admission. Chrome PID 6128 was in ptrace_stop.part.0, state t, with zero user/system ticks in all 101 samples. The monotonic sample window was 10164.210123 ms. Twenty-two synchronous journal writes overlapped 10125.790473 ms of it; nested fsync calls overlapped 10117.760935 ms. These totals overlap each other and must not be added. Kernel records show real fork/exec/signal stops and releases for that launcher. This identifies tracer durable-service delay at the launcher boundary; it does not establish a Chromium internal fault, device-only latency, or guest failure. The browser was alive but repeatedly stopped, not independently making progress.

Nox now generates custody-run.py, reusing the published Q3 release-cohort adapter from 7607cab84. Its authenticated load() supplies the shared kernel supervisor; Q3's app-specific observation wrapper is not used. Both Python inputs are exact size/SHA pins in the reviewed full closure and are checked before loading. The Q3 adapter and shared custody provider are unchanged here. The native caller must use this generated entrypoint. Browser readiness remains 100 polls of 100 ms; protocol, runtime and cleanup deadlines are unchanged.

The actual generated consumer was exercised under real ptrace with a Node browser peer, twenty real fork/exec children, delayed port publication and 80 ms synchronous storage service. With the old Nox adapter it exhausted all 100 polls, admitted no protocol and reaped 38/38 tasks. The reused service published readiness at poll 56, completed ordinary shutdown and reaped 38/38. Fsync count fell from 170 to 95. Missing-port and malformed-port controls remained terminal with zero protocol, actual child waits, separate stdout and stderr EOFs, full kernel drains and unchanged deadlines. Adapter/provider corruption controls refused before load. This is a causal offline service control, not a claim of native Chrome or compositor acceptance.

Readiness repair bce2b316f was published before the sole native attempt. The generated fixture publication, foreign-owner and changed-image controls and twelve observer command/image bindings passed, as did the locked build. A second local plan inspection initially refused existing provenance patches with EEXIST. Using a fresh plan directory resolved that packaging error before allocation; both the error and passing result are retained. The native plan pins the exact published readiness commit, and its inputs match the tested offline closure byte for byte. No source archive or repeated media copy was made.

Native readiness succeeded: four polls, 460.687499 ms in the monotonic sample window, followed by Browser.getVersion and the final app-selection boundary. The exact first fixture PNG was separately model-viewed and hash-checked. It shows the dark fixture panel, title, green TARGET and initial yellow crosshair. The review authorized one native center click with independent 150 ms release.

The first live failure was the fixture's document-level DOM count assertion 1 != 2, before its after-image. The owning supervisor received immediate pidfd SIGTERM after the terminal record was observed. No later game protocol, input, capture, retry, quit or deadline extension followed. The conditional original was not started; Jack, Bunker and staff impact remain untested.

Retrieved records prove renderer handleMouseDown and handleMouseUp at logical (320,240), matching DirectInput delivery types 1 and 2, tail 2 and final buttons zero. The document observer saw trusted mousemove and mousedown, but no mouseup. The source-backed cause is separate from readiness: browser-input.js installs windowMouseUp at window capture after mousedown; it stops propagation before the observer's document capture listener. The bridge successfully delivered release. This is a fixture observation failure, not guest nonresponse or lost native release. Exact source pins and the native facts are retained in native-observation-cause.json.

The same worker reproduced this with the actual generated consumer by correcting the offline peer's capture order to window → document → target and honoring stopPropagation. Previously the peer delivered document events first, hiding the bug. The fixture-only observer adapter now routes mouse listeners to window capture before the bridge arms drag capture; pointer-lock events and getters remain on document. It still requires both trusted down/up events. The shared observer, production input bridge, Q3 adapter and custody provider are unchanged. All four final generated controls passed: positive, foreign owner, changed PNG and dropped release. Every negative remained terminal before the after-image, with actual child waits and separate EOFs. The final locked build passed. This second repair is tested offline only: the single native attempt had already ended, and no additional native attempt was made.

Cleanup reaped all 151 owned tasks with complete kernel fork lineage, actual parent wait 1, separate stdout/stderr EOFs, no surviving named processes, closed ports and absent profile. All 102 artifact references (99 unique contents, 1543280 unique bytes) were retrieved and independently reread by SHA-256 and size, with gzip/PNG validation, before the sole normal unforced stop.

Box bx_k96mm34m is archived, expiry null, independently confirmed with stop receipt stop_12fe81b31d1f856b, completed unforced at 2026-11-10T01:32:07.104Z. Prearmed watchdog UUID 3e7c40a6-1bcf-48f2-bdfb-7bb414831d55 retained its immutable deadline 2026-11-10T01:54:51.093Z; initial API expiry was 01:54:52.434Z. Its independent supervisor recorded actual watchdog wait 0 and separate EOFs. Home remained above 1.5 GB free. No owned live box remains; bx_xk729gjr and all other prior boxes were untouched.

Next is a separately authorized native fixture for the window-observation repair, then the conditional Jack/Bunker staff route if that fixture passes.