VB5 CCE: nested owned MessageBox and supervisor shutdown

Evidence: scratch/runs/20261117-vb5-nested-owned-modal-install/.

The preceding original pipeline remains permanently closed. Its retained public snapshot showed MessageBox 0x10018 owned by live DialogBox 0x10012, with no main window. The previous scene validator refused modal.ownerless; the owner was not zero. Its outer exit 143 also lost native/watchdog wait and descriptor release records. Process absence established cleanup observation only.

The native MessageBox API now seals its creation against the active DialogBox invocation. Actor-local native values retain both HWND generations, the owning process/thread, the DialogBox callback and return site, and the MessageBox call return site. The observation export independently checks the live common-modal pump, actual USER class/owner/thread, original DialogBox invocation, callback, generations and return sites. Destruction, retirement, reuse, actor changes or callback replacement invalidate the seal. No title or application-specific identity participates. Scene admission requires this chain to match the actual public owner and MessageBox identities. Class 16 and a live owner alone confer no new admission.

An authored PE calls MessageBoxA with YES/NO + QUESTION from its real DLGPROC, then returns to the exact outer invocation. Real native BUTTON inputs exercise the transition and subsequent main creation. Fault cases retain the existing finite-input cancellation and permanent UNKNOWN behavior. A separate real CPU Worker runs the same PE and reads the native seal through its actual export channel, including owner/generation/actor/callback/retirement refusals and return restoration. This does not claim that the CLI's main-instance focus snapshot can capture an arbitrary Worker-owned GUI; that separate limitation remains fail-closed.

The Python supervisor catches outer TERM/INT, preserves exit 128+signal, and unwinds through native process-group termination, actual wait, watchdog pipe drain/join and sealed descriptor release. Repeated shutdown signals cannot interrupt that cleanup. Regular native log files are described as regular files, without invented pipe EOF claims. Actual native/watchdog/outer fixtures cover normal exit, TERM143, exceptions, timeout, native EOF and watchdog deadline.

Original media and the genuine same-package SetupAPI/CFGMGR32 staging remain unchanged. Original execution now seals one exact published Git tree, including all helpers; local overlays are allowed only for authored working qualification and are rejected for any original launch. The final result records publication, qualification, the one authorized fresh original outcome and its next boundary.

Published qualification and original result

Material commit: efcf47369966a1bcc1f0e8b0a2d74f99c1eca293, published on main with strict pull/push exit 0. The required canonical flock build passed. Fresh qualification and the single original used that entire published Git tree, without overlays, with module SHA-256 2c7d23d9c735bdec75884528482e1b1d8c73334bd02c8d1f3b15850011b57093. The empty owned-helper patch and excluded worktree patch are retained; no older native-core composition or shared WIP entered the original.

That exact module passed the nested native suite (positive plus nine refusal cases), six actual CPU Worker cases, six supervisor shutdown cases including truthful outer exit 143, and all existing DialogBox, ownerless, QUESTION, controller, progress, mapping, context and button guard suites. Genuine SetupAPI/CFGMGR tests passed in main, Worker and child execution, including real image exports, INF/queue operations, retirement and sealed inheritance. Each subprocess has its actual wait record. Earlier working-module results are not substituted for these fresh published-source proofs.

The fresh NT4/defaults/real-tick original passed the former boundary. The public snapshot and native seal agree on MessageBox 0x10018, generation 24, owned by live DialogBox 0x10012, generation 18, process 1000/thread 1. The outer callback is 0x1ab9a44, outer return site 0x1ab9a3e, and MessageBox return site 0x1ab99b7. Its personally reviewed create-folder Yes completed in 210 ms with independent physical UP. The installer then displayed installation complete. The flushed overlay has 258 records, including the 1,818,384-byte C:\Program Files\VB5CCE\VB5CCE.EXE and seven sample .vbp files; the registry export has 836 entries. This establishes installation output, not IDE or sample execution. The native log also retains the unavailable grpconv.exe diagnostic and held-critical-section-at-exit warning.

Final installation-complete OK had native DOWN CONSUMED and independent physical UP ACK. Its mandatory post-click snapshot failed with connection refused as the installer exited, so the controller wrote PERMANENT UNKNOWN. There were no later HTTP requests, rebindings, stage-one launch, replay or second original. IDE/sample/nativeQuit remain unattempted. The concrete next boundary is reviewed-intent completion across a normal native process exit: it needs an authored fixture and truthful native wait integration, preserving fail-closed UNKNOWN for ambiguous exits. A closed endpoint alone is not successful completion.

The original native process was actually waited with exit 0; its output was in regular files, with no pipe-EOF claim. The independent watchdog was actually joined with exit 0 and both pipe EOFs. The supervisor tool session actually returned 0, and runtime-release.json records all owned descriptors closed. The TERM fixture separately retains actual outer exit 143 and all native and watchdog joins. Process/port absence is corroboration only. Eight selected named images remain, with hashes; redundant qualification/extraction images were removed. No box was created: ID null, state NOT_CREATED, expiry null, stop receipt not applicable. The old original pipeline remains closed.