Caesar III fresh reviewed native Quit
Evidence: scratch/runs/20261117-caesar3-fresh-reviewed-native-quit/result.json.
The actual external CLI sequence has been proved on an authored guest-drawn File/Exit menu, matching Caesar's custom menu USER path, followed by a genuine owned native MessageBox. Each image was opened in full and its current public capture metadata inspected before one immutable review. File and Exit used native USER-return receipts; Yes used the published physical native BUTTON path without a separate MOVE, with native-button-dispatch-return, independently prearmed physical UP, and actual normal native exit 0. The Yes operation took 473 ms, below the unchanged 500/1750/3000 ms bounds. Actor, focus, modal, generation, nonce, target and 15/30-second freshness checks remain unchanged.
A first authored harness had an evidence-directory environment error before guest launch; it was closed with actual waits. A separate standard Win32-menu fixture correctly closed UNKNOWN when tracking produced no native input edge. That fixture does not represent Caesar's guest-drawn menu. Neither closure was retried on the same endpoint and neither authorizes an original-game fault claim. No core or generic input change was made.
The adapter now accepts the explicitly selected CAESAR3_NATIVE_SECONDS=600
with independent watchdog 620 seconds and Quit reserve 180 seconds, while its
existing default remains 1200/1220. Other durations are refused. Every adapter
function and the published modal controller remains byte-identical to 99fe9e041;
the executed c993/aa25/df365 module remains SHA-256
9099bb3f694d5ec6dac2886b70c1920fa33f39ad9258b1e662d1ebe05dc2da0a.
Matching completed component and native guard proofs are reused by exact bytes.
The manual CLI fixture entry point retains supervisor wait and EOF evidence.
The one fresh original used the published 397ad979f adapter with ordinary 600-second native / 620-second watchdog / 180-second reserve limits. The minimal route reached an untouched city (Pop 0), then File > Exit. No road, construction or population work was repeated. All earlier native USER-return / independent UP operations completed inside the existing bounds.
The original created native Quit 0x10002, generation 2, owned by main 0x10001,
and Yes BUTTON 0x10004, generation 4, parent 0x10002, control ID 6. But the full
capture still showed the city/File menu. A second ordinary fresh capture 12.74
seconds later had exactly the same PNG SHA-256
2f3b8d0c01c965b42b41ed8155f80a5be463bbde33f8ce64c729cf7efd910f3b
with the same public Quit/Yes identity. Neither image supported a Yes review.
No Yes intent or DOWN was sent. The endpoint was permanently closed UNKNOWN;
there were zero later HTTP requests. Native status 0 was host cleanup, not
a normal native Quit. No second original was launched.
An authored guest using public DirectDraw COM calls reproduced the missing owned MessageBox on the exact 9099 module: complete primary Unlock, then real MessageBoxA. The delayed host primary upload stamped a newer coverage order than the dialog's already completed native GDI paint, incorrectly hiding it. The host now records the primary presentation order at the existing native PRESENT trace and consumes that order once during upload. This changes no guest pump, paint, callback, input, clock, module or receipt. Later real covering presents and retained CPU uploads without pending presentation still cover older GDI; clipped coverage and normal repaint remain unchanged. Pending order is removed when its native surface retires.
With only the leased host presentation source overlaid on the sealed graph, the same authored PE/module shows the real Quit, Yes and No over the primary. This is a rendering-only diagnostic and was closed by explicit host cleanup; it is not a native Quit claim. The upload/compositor regression fails on the old host and passes after the fix. Existing dialog freshness, forced-pixel and exclusive no-synthetic-paint tests pass. The first canonical build correctly refused the changed host producer hash; its scoped lifetime review was updated before the required locked build was repeated.
The original native Quit remains unproved. The spent endpoint must not be queried or reused. Any later original requires a new explicit assignment and must first qualify the unchanged modal helper's strict pointer-mapping contract: the original observed an identity exclusive transform, whereas the current helper requires no transform. That guard was not relaxed in this task.
No box was created (ID/state/expiry: null/NOT_CREATED/null; stop not applicable). All created native, adapter, outer, owner and independent watchdog processes have actual wait and both-pipe EOF evidence in the run. Media remained unchanged.