VB5 CCE first-dialog context, 2026-11-17
Evidence: scratch/runs/20261117-vb5-first-dialog-context/.
The Nov16 launcher has no final join receipt. Its recorded compile PIDs are
absent; no cleanup signal or request to an inherited endpoint was sent.
Missing launcher provenance remains UNKNOWN. The prior original refusal in
199ed3c6e does not identify the failed predicate.
The continued diagnostic change reports bounded field names, separating native context predicates from unsupported class/unavailable text predicates. It emits no captions, text bytes, paths or private application state. It does not expand capture admission. Generated NT4 install-question fixtures exercise three actual USER shapes: no main with NULL MessageBox owner, main-owned, and NULL owner with an existing main. All use actual child BUTTON focus/labels. The first shape refuses main-dependent checks; both shapes with a main capture successfully. The two complete successful generated images were personally viewed, after testing, without claiming original review or input.
The no-main shape is a demonstrated candidate defect, not yet the established cause of the original refusal. One separately authorized passive original attempt may determine that cause after diagnostic publication. It permits no binding or input, uses fixed 120 s native / 140 s independent watchdog bounds, and closes on its first refusal. The interactive original authorization remains conditional on causal repair, native qualification and publication.
The runtime uses the published indexed source loader and source/module seals for parent, actual Node Worker and spawned Node inheritance. Only enumerated owned helper patches overlay one Git tree; other workers' WIP is excluded. The shared canonical build is separate from that immutable runtime.
Canonical build passed after Boids' owner repaired its renderer lifetime pin;
the intervening failures are retained. Focused field/context units and the
three native shapes qualify the passive diagnostic only. Optional wider
qualification attempts exposed two incompatible legacy harness routes (loader
replacement and explicit Worker execArgv replacement). Indexed inheritance
and the two real desktop receiver tests pass; the standalone desktop observation
suite also passes. A later generated positive Quit exited before its independent
UP request could connect, so that full interactive qualification is a failure,
not admission. Its UNKNOWN and actual teardown evidence are retained. Interactive
original execution still requires successful qualification after causal repair.
Diagnostic publication 0856ec434 preceded the one passive original. Its
independent prelaunch audit verified all sealed sources/module and both genuine
media hashes. The first and only HTTP request failed with main.windowPresent,
main.nonNullHwnd, the other main identity fields and their dependent equality
predicates; controls was empty. This establishes the actual absent-main cause,
matching the generated NULL-owner/no-main fixture, rather than assuming it.
No PNG or input was produced. Permanent UNKNOWN closed the endpoint; native
actual wait was 0 after host SIGTERM, watchdog actual wait 0 with both EOFs, and
the supervisor closed its descriptors. This is cleanup, not ordinary guest Quit.
The causal repair admits absent main only with a live, visible top-level native MessageBox whose actual native record matches HWND/generation/process/thread, whose owner is NULL, and whose focused real control belongs to the same actor. A missing nonzero main, arbitrary ownerless window, foreign thread/process, missing native box, stale generation and unsupported controls still refuse. No USER parent/main metadata is changed. The permanent input binding for this case pins the real registered guest actor, renderer/device/queue and OS process; actual HWNDs and generations remain pinned by each retained reviewed intent. Main-based bindings retain their previous checks.
Generated qualification now includes a pre-main MessageBox followed by ordinary guest main creation and two further modal generations under one binding. The authored test guest sleeps for one real second before ExitProcess to allow independent UP to finish; no original EXE or clock is changed. This explicitly addresses the earlier fixture's immediate-retirement race without counting it as successful UP evidence.
repair-qualified-all passes all 13 NT4 native cases: normal and repaint
positives, the three-generation pre-main/main-creation positive, and five faults
each with and without an initial main (focus, renderer owner, modal caption,
real child BUTTON caption, and actual elapsed stale review). All fault cases
retain permanent UNKNOWN and assert zero later HTTP, including local replay
attempts. The three complete pre-main/main-created/second-dialog images were
personally viewed after qualification. Units also refuse missing native-box
records, foreign identities, retired generations, disabled/changed targets,
missing nonzero main and unavailable text/classes. The final canonical flock
build passed. Interactive original admission remains a separate finite launch
after publication, using exactly the qualified source hashes.
Causal publication f6682173b preceded interactive preparation. The first
preflight passed the three exact question capture/binding shapes but refused
before any original launch: a fixture native-wait JSON was visible while being
written, and the controller correctly latched UNKNOWN on its incomplete body.
The writer now fsyncs a complete pending receipt and links it into place without
overwriting. Original supervisor completion/launch receipts use the same atomic
publication rule. The failed preflight's originalStarted:false is retained;
it consumed no interactive original attempt and supplies no success claim.
Bounded original outcome
Atomic publication 9b723227c passed the exact three question capture/binding
shapes, all 13 native cases, units and canonical flock build. The final holder
used Git f6682173b plus the two enumerated atomic-receipt patches, both backed
by published 9b723227c. An independent audit rehashed all sealed sources,
module, currently executed owned helpers against that publication, and genuine
installer/SetupAPI media before opening the sole interactive gate. Fixed
2400/2420 s native/watchdog deadlines and 240 s reserve began after preparation.
The first original complete 640x480 image was captured and personally viewed.
It shows the install question with focused Yes. The actual target was native
BUTTON 0x10003, generation 3, control ID 6, parent modal 0x10001, with no
main HWND. One finite click at (209,139) completed in 230 ms with independently
prearmed physical UP and durable evidence within the original limits. This is
actual original proof of the exact first-dialog repair, not just a fixture.
The second capture, provisionally named install-directory, refused on the
five messageBox.* identity predicates with controls:[]; no second image was
written. The retained post-click public USER snapshot identifies license
controls under a new live, ownerless top-level modal 0x10005, generation 5,
control class 0, with mainHwnd:0 and nativeMessageBox:null. Its focused EDIT
is a real child of that modal. This is not a personally viewed license screen,
and no agreement acceptance occurred. It demonstrates the next trust-contract
gap: an ownerless pre-main modal without a native MessageBox identity record.
It does not justify whitelisting arbitrary class-0 ownerless windows. A future
repair needs a separately proven native dialog identity and native equivalents.
Permanent UNKNOWN closed the entire pipeline on this refusal. There were six
interactive endpoint requests (capture, bind, input, independent release,
completion snapshot, refused capture), exactly one finite UI input and zero
later requests/replay/rebind. Host SIGTERM of the exact recorded process group
produced native actual wait 0; that is not guest Quit. Watchdog actual wait was
0 with both pipe EOFs, no deadline firing, and supervisor descriptors closed.
Installation, IDE, shipped samples/axlstpik/testgrp.vbg execution and ordinary
guest Quit remain unproved. CCE standalone EXE compilation is not claimed.
Postflight finds all 487 recorded task PIDs absent and both attempts' four ports without listeners. Both media hashes still match. Six named selected images remain (one original question and five generated native proofs); other generated images were hashed before removal. Git plus explicit binary patches is the source archive; no source/media/WASM archive or capture bundle was made. No boxes/workers were created: box ID/expiry null, state NOT_CREATED, stop receipt N/A. The missing Nov16 launcher joins remain UNKNOWN. Both newly used original endpoints are permanently closed; this result authorizes no replay.