Delphi native controller completion and capture qualification

2026-11-16. This assignment runs generated native USER fixtures only. No original installer, installation, credentials, IDE, samples or original Quit is claimed. Evidence: scratch/runs/20261116-delphi5-native-controller-completion-repair/result.json.

The preceding controller slept until 2500 ms before beginning its observation, leaving only 500 ms for checks, durable records and scheduling inside the fixed 3000 ms aggregate. It now observes the actual modal transition immediately after the independently armed physical release. Quit waits for the parent's actual native process completion and pipe drain, without requesting a retiring endpoint. A diagnostic preparation exposed and retained that exit/socket ownership race. The 500 ms confirmation, 1750 ms physical UP and 3000 ms aggregate limits remain unchanged. Physical UP timestamps now record the returned acknowledgement, rather than the time immediately before sending its request. Freshness remains 5 seconds for gameplay and 30 seconds for menu/install/Quit evidence, with the consumption deadline also capped by evidence expiry.

The final run recorded the following actual times, including durable evidence publication before the completion measurement and final aggregate assertion:

Native intent Confirmation Physical UP ACK Aggregate
First modal OK 244 ms 392 ms 448 ms
Second modal OK / native Quit 255 ms 377 ms 523 ms

Another preparation exposed Python interpreter startup exceeding the existing 150 ms release-prearm check before any DOWN. The release helper now forks from the single-threaded controller into a separate process, avoiding a second interpreter import/startup. Its prearm check, absolute 1400 ms fallback and all three immutable action limits are unchanged. The parent actually joins that process and drains both pipes before accepting its physical UP evidence.

The old fixture passed the same title and text to both MessageBoxA calls. Its identical pictures cannot establish whether the second capture was stale. Contrary to the preceding report's description, the retained old second frame shows a generated dialog, not a blank teal image. This repair gives each real guest-created dialog distinct strings and checks its native generation, title, full-frame dimensions, rendered content and different PNG hash. Capture records also compare native context before and after PNG publication. Personal review of the two complete frames is recorded separately from automated fixture reviews.

Native scene identity alone did not guard paint changing while an intent waited. The CLI now accepts an optional complete-frame SHA-256 inside a native intent and checks it during both preparation and consumption. This controller supplies it on every intent. Existing native clients retain their protocol; this qualification requires the stronger paint guard. The published desktop ready-loop repair c8f308b84 is preserved without edits.

The historical post-UNKNOWN generic fixture teardown request is removed. All controller requests, including cleanup, check the permanent terminal latch before identity checks or network access. A failed action drains its already armed finite release before publishing UNKNOWN; no release can bypass the published latch. Fixture errors also latch permanently. Remaining live fixture processes are terminated by their owned host process group, never by a teardown endpoint request. Host termination is explicitly distinct from native guest Quit.

A later preparation also refused before capture/input when the previous fixed fixture port belonged to another worker. It sent no request to that listener. Fixtures now obtain an OS-allocated free local port once, and retain the exact PID/start-token/socket ownership check against an intervening bind. There is no automatic replacement or endpoint retry after a refusal.

Qualification includes actual queued focus, owner and paint changes with unconsumed cancellation; stale paint at binding and at a later intent; and capture/bind/retry attempts after UNKNOWN with unchanged request ledgers. Independent deadline watchdogs are armed before native process launch. Native, release helper, watchdog and supervising checks retain actual waits and pipe EOF evidence. No box, shared build-output reader, source archive, module dump, foreign WIP, installer media mutation or clock override is used.

The runtime helper now defaults to qualification and closes its sealed source and module descriptors on success or failure. An original pipeline requires the separate explicit --original option and its existing review gate in a subsequent assignment. Git plus the owned binary patch and new-file patches reconstructs the tested sources; the canonical compiler builds their module directly into sealed owned memory.

Final checks and publication are recorded in the evidence result. The v2 preparation retained a desktop regression timing refusal at 3424 ms against its unchanged 3300 ms test ceiling; it launched no native fixture or original. That failure is not rewritten into success by later passing runs.

Final evidence is in final-qualification-v8/. The old Git host's real queued paint fault consumed stale input; its expected-red trace is in paint-baseline/. The repaired host cancelled the queued paint, focus and owner faults without consumption, and rejected both stale-paint binding and later-intent submission. All five permanent UNKNOWN cases recorded zero later requests. Three selected full-frame screenshots were personally reviewed: the first and second native dialogs, and the deliberately changed paint used by the rejected intent.

The required full-flock build passed after the native CLI and fixture changes. Only the Python release-process startup implementation changed afterward; the final focused runtime tests cover that change. The compiler/build inputs and native CLI/fixture bytes remained unchanged. A preparation that hit its unchanged 180-second compile watchdog is also retained, with actual TERM wait and both pipe EOFs; it launched no native fixture or original.

Next: a separately assigned original trial installation using the qualified controller, ordinary original media and fresh personally reviewed scenes.