Delphi 5 after the desktop repair: native qualification refused

2026-11-16. No new original installer was launched. The required native qualification failed its fixed 3000 ms aggregate guard, so installation, credentials, IDE execution, samples and original native Quit remain unproved. The previous original endpoint was never contacted, rebound or replayed.

Evidence: scratch/runs/20261116-delphi5-trial-after-desktop-repair/result.json. This report records a controller/host qualification blocker, not a Delphi API failure or installation result.

Exact inputs and successful checks

The final preparation used published Git 6f6289091fdfeab5de8868b9d2d60506386a6af1, including desktop repair c8f308b84d73a62a5aaf9e8c8cf908fb766b8346. Its runtime JS came from that Git tree, with the owned test helper overlay recorded as a binary patch. Foreign working-tree changes were excluded. An earlier board entry naming 915b0036d applies to earlier preparations; the final source manifest is explicitly pinned to 6f6289091.

The vendored compiler compiled the Git source closure directly into an owned memory descriptor, using canonical tail calls and the region-layout stamp. Module SHA-256 was 61a7d750c83c62ecbbe6e2388136d6a34d152f5782cb58b0338aea2a8528cc2b (2,277,621 bytes). Runtime sources and the module were sealed and held by this worker's supervisor. There was no source archive, WASM dump, borrowed holder, shared build-output mutation or box. Git, the source hashes and owned patch provide recovery.

The untouched ISO's 81,330,176 bytes and SHA-256 bcd14e38635f4b3b4c7656f1ce884c26ba60224d7876681a7e3b7382e5c38b0e were verified, and hashes of its 38 extracted files were recorded. No payload was substituted, cabinet entry launched separately, key invented, provider faked, trial clock changed or guest state modified. No new guest VFS or registry was created because the original launch gate never opened.

The real Node Worker baseline returned the preceding reply with the old receive implementation. The published repair passed all three focused tests, including repeated unready notifications under the unchanged deadline. The existing real desktop regression passed its socket, lost-reply, authority-death and terminal no-retry cases. Native scene unit guards also passed. These checks do not establish that the complete controller qualified.

First final budget refusal

The generated PE fixture used ordinary native USER dialogs, with two distinct modal generations. Its automated reviews explicitly made no personal original review claim. The final measurements were:

Intent Confirmation Physical UP ACK
First dialog OK 103 ms 201 ms
Final dialog OK / Quit 115 ms 210 ms

The generated process exited 0 and was actually waited. The controller recorded that wait at 2501 ms without a post-exit endpoint request. After publishing its observation record, its aggregate completion assertion failed; its permanent UNKNOWN record was written at 3069 ms. The interval includes local evidence publication/fsync and scheduling; their individual causal contributions were not measured. This is not evidence that native Quit took 3069 ms, nor that the desktop repair failed.

The fixed 500/1750/3000 ms limits were retained. No new endpoint or original was launched after this refusal. The final focus-fault fixture was NOT_RUN because the positive path failed first. A previous preparation passed the positive path but failed to supply the required PNG hash to its focus-fault bind; that is not a passing fault test.

Only these final fixture screenshots were personally inspected:

Preparations, waits and release

Six earlier preparation directories retain failures rather than rewriting them into success. They include a killed oversized source-view regression, a missing observation directory, post-exit observation attempted before the fixture parent could process its exit event, and a schedule-file read/write race. The prepared helpers now use atomic record publication and asynchronous fixture-parent waits, but remain unqualified WIP and are not committed here.

The first detached preparation supervisor has no recovered parent wait: its outcome remains UNKNOWN. Its logs are regular files, not pipe EOF proof. Later invoking tools reported exit 1 separately from the recorded native and watchdog pipe waits. Final native PID 3406135 and watchdog PID 3406128 both have actual exit-0 waits and both pipe EOFs; final qualification child exit 1 and its supervisor's tool exit 1 remain distinct.

There is also a historical protocol defect in preparation v4: after its helper marked UNKNOWN on the schedule-file failure, the fixture parent sent its generic teardown quit request. That preparation is unqualified and its terminal zero-later-request claim must not be generalized. In the final preparation the native process had already exited before UNKNOWN, so no teardown request followed. No original endpoint was involved in any of these preparations.

All recorded task runtime PIDs are absent, ports 45761/45762/45861/45862 have no listeners, and process exit closed the owned descriptors. Box ID/state/expiry and stop receipt are inapplicable: no box was created. No shared build window was taken. The helper build request was withdrawn when qualification failed; only this report is committed, so no code-build success is claimed.

Next work must qualify the controller's complete fixed-budget path, actual fault rejection and visibly rendered fixture scenes before a separately assigned original installation. It must also fix the fixture teardown's historical post-UNKNOWN request path. The known HEAP_SHARED/global-sync contract remains unchanged and must not be bypassed in any original run.