VB5 CCE reviewed-context integration, 2026-11-16

The single authorized original run stopped before the first input. Its first native-reviewed-capture returned native capture context unavailable or unsupported semantic control class and latched permanent UNKNOWN. Exactly one endpoint request was made. No original image was written, and no recapture, rebind or second original occurred. A later local intent refused before any network request. The precise rejected context/control field is unknown because the refusal did not include a field-level comparison and the closed endpoint was never queried again.

This run integrates native-reviewed-context-v1 for the original VB5 CCE installer and the intended installed samples/axlstpik/testgrp.vbg. Evidence is in scratch/runs/20261116-vb5cce-reviewed-context-install-ide/. Installation, IDE/sample interaction and ordinary guest Quit remain unproved. The native log reports the initial install question, but there is no current original screenshot or claim of personal visual confirmation of it. Host SIGTERM cleanup returned native exit 0, which is not ordinary guest Quit.

Native child-control defect and repair

The first NT4 generated fixture exposed a prerequisite defect in the published contract. controlSnapshot.windows enumerated renderer.windows, but native MessageBox controls share their parent's backing canvas and have no renderer window. The snapshot listed main 0x10001 and modal 0x10002, while its actual live focused BUTTON was 0x10004. Binding that actual target returned reviewed target unavailable; the controller latched permanent UNKNOWN and issued zero later requests. A modal-container pointer target could consume an input but did not qualify the real button's semantics for installation.

The coordinator authorized a narrow repair to released core files in the same task. lib/cli-native-user-windows.js now walks public native USER child relationships and includes their HWND, generation, process/thread, parent, owner, style, effective visibility, geometry, control ID, class and caption in every capture and consume-time snapshot. Pure read-only exports return the same native BUTTON/EDIT/STATIC text that WM_GETTEXT exposes. They allocate nothing, send no messages and read no application-private state. Native button captions are stored in ButtonState, not the ordinary window title table. EDIT snapshots also include public selection/cursor/flags; password controls remain unsupported. Other unsupported control classes still refuse capture. The build's control-variant attribution documents the class-2 EditState branch.

The VB-only controller retains the complete original reviewed image and its capture times, identifies one exact finite target, and uses the retained-image contract at binding and consumption. It no longer substitutes a new capture for the reviewed image or requires identical repaint bytes. Process/socket identity, public USER context, 5 s game/30 s menu freshness, 500 ms confirmation, 1750 ms independent physical UP and 3000 ms durable completion remain enforced. Every refusal/UNKNOWN closes subsequent requests permanently. Generated fault injection is confined to authored fixtures, never original media.

Qualification and original admission

The repaired pipeline-qualified run passed all seven native cases. Every native log reports NT4 version 0x05650004. Positive and repaint-positive each completed both modal generations and ordinary ExitProcess with actual parent waits and both pipe EOFs. Targets were actual native BUTTONs 0x10004 and 0x10007, with public caption OK, rather than modal surrogates. Confirmation was 68.623–75.498 ms, independent physical UP 243.477–317.370 ms, and durable completion 287.534–420.836 ms. Both complete positive images were personally viewed after the automated generated-fixture run. This is diagnostic fixture evidence, not personal review before original input.

The tested immutable module SHA-256 is 5c9f7126daff2e586685c0dd2ad46b79c5e43e8dbf569aea2919b733b075614a. The exact source Git identity, selected file hashes and owned binary patches are retained with that runtime. Material commit 431276180 was published with strict pull/push before the sole original gate opened. The plain canonical flock /tmp/wine-build.lock bash tools/build.sh passed with actual tool wait 0, as did the guard, input-owner and control-variant regressions. A preceding 300 s build wrapper expired while compiling after including the foreign lock wait; its actual 124 is retained separately and is not counted as a pass.

The NT4 fixture uses two distinct guest-created native MessageBoxes and ordinary button interaction/ExitProcess. The positive repaint variant changes complete PNG bytes while keeping the same native target and semantics. Negative cases change focus, renderer owner, modal caption and actual native BUTTON caption between preparation and consumption; stale menu evidence uses real elapsed time. Closure checks attempt local later capture/bind/input and assert no new endpoint requests. These generated protocol tests use explicit fixture attestations; they do not claim personal review of original installer frames.

Earlier preparations are preserved honestly: qualification 1 completed positive and repaint transitions but failed on a duplicate result filename; qualification 2's supervisor returned 143 after compilation with no normal release receipt. The optional qualification-final personal fixture run completed four reviewed diagnostic inputs against modal-container targets, then failed an added assertion that the ordinary window title table contained the actual BUTTON caption. The repaired child-census-1 and pipeline personal fixture gates expired before input. The pre-repair qualification-3 completed its diagnostic suite and recorded original admission as false. These were generated fixtures; none launched original media or contacted an old original endpoint.

The immutable runtime is compiled from one recorded Git tree plus explicit owned patches, then sealed in memory. Other workers' uncommitted source changes are excluded. No source tree, module, media or screenshot bundle is archived. The canonical shared build is a separate required check and is not the runtime module. Shared build failures from the initial missing EditState attribution and concurrent host/renderer lifetime pins are retained, not counted as passes.

The sole authorized original pipeline verifies installer SHA-256 022af94b651cd9d31e0676a6819dd5085940893f1ecc4abaa85b48c30de6f318 and genuine SetupAPI SHA-256 51f2682279e28386aa9c0eb8ba4f8ac3000c9ff75cf7e2456e8cdd9f0c544159. The retained DirectX 7 package and extraction recipe are documented in vb5cce-current-original-20261116.md. Media/defaults remain genuine. Runtime and children declare NT4 and use real clocks. Preparation completes before the fixed 2400 s native deadline, independent 2420 s watchdog and 240 s Quit reserve begin. Original inputs require personal review of a complete current frame. No unattended UI action sequence or second original is authorized. Before launch, a separate audit reopened the supervisor's source/module descriptors, verified both seal masks were 15, verified the exact module hash and every selected source hash, and rechecked the mutable VB controller/test hashes.

The original used the verified media, declared NT4 version 0x05650004 and real clocks. It launched no setup/IDE stage and received no input. Its native actual wait was 0 after SIGTERM; regular stdout/stderr files closed, so pipe EOF is not claimed for that process. Supervisor and independent watchdog actual waits were 0, the watchdog's two pipes reached EOF, and the supervisor closed its owned sealed descriptors. Deadline/watchdog signals did not fire. These process results establish cleanup, not installation success.

CCE runs its shipped sample inside the IDE; no standalone EXE compiler capability is claimed. No boxes or additional workers were created: box ID and expiry null, state NOT_CREATED, stop receipt N/A. Telegram progress/final updates were delivered as messages 2900, 2902 and 2903. Keyed replies to inbox 200/201 already had different saved replies; those guards were respected and the updates were sent as follow-ups naming both requests.

The next blocker is the original first dialog's rejected native context/control predicate. Diagnose it using a generated equivalent and field-level refusal evidence before seeking separately authorized original execution. This task's original endpoint is permanently closed, and this result authorizes no replay.