Drakan retired modal focus — 2026-11-18
Evidence: scratch/runs/20261118-drakan-retired-modal-focus-gameplay/ (ignored).
The two predecessor originals and their UNKNOWN scopes remain closed.
An ordinary authored PE reproduces the real pre-main NULL-owner MessageBox,
native Enter consumption and independent UP, followed by RegisterClassA,
CreateWindowExA with the measured hidden 0x00ca0000 style, ShowWindow(5),
and an ordinary message loop. Guest GetFocus immediately after MessageBox
returns the retired IDOK handle on the predecessor module. The guest exposes
that API result through its main-window title, independently of the observer's
focus field. Normal activation subsequently gives this authored main valid
focus; the retirement defect is demonstrated, but is not an exclusive explanation
of every authentic post-warning failure.
wnd_table_remove now releases precisely the retired window's local focus
before host retirement callbacks. Native modal tree teardown previously bypassed
the clearing in wnd_destroy_recursive. Unrelated focus is preserved; ordinary
modal restoration and ShowWindow activation retain responsibility for selecting
the next focus. GetFocus validates liveness and the calling queue's thread and
returns NULL for a stale or foreign handle, without assigning another window.
The quiet-handler inventory decreases by one with this implementation.
This follows the official GetFocus contract and DestroyWindow contract. No Windows reference run or Wine was needed.
The corrected module passes the same authored transition: guest GetFocus is NULL after warning destruction and normal activation selects live main focus. A real worker_threads actor and idle instance over shared WND memory demonstrate the distinction: the owner clears its focus while the shadow may retain a raw private handle; GetFocus cannot expose that retired handle. Native controls cover unrelated live/hidden focus, destroyed owner, calling-thread boundaries and hide/noactivate commands. Existing activation, owned-form ShowWindow, window state, renderer modal routing, and nested DialogBox/MessageBox continuation checks pass. A real-tick 500 ms transition retains every pending response and its full failure list in append-only JSONL, under the original three-second deadline.
Diagnostic limitations: the historical browser-worker focus test stops at a source-text assertion about process routing; the older modal-worker test expects idle yield 15 but observes 0 before reaching teardown. These are not reported as passes. One authored 100 ms pulse closed on a late consumed ACK despite physical UP acknowledgement; a separate authored control found exclusive receipt-name collision, fixed by append-only JSONL. Those scopes were cancelled and actually waited, without later endpoints. Fresh authored controls pass with a finite 300 ms pulse and unchanged 500/1750/3000 ms bounds.
The authentic run uses the predecessor's proven source graph and genuine media closure, with the published modal-key and focus corrections. Git, owned patches, the sealed compiler recipe and unique module provide recovery; no source-tree snapshot or media copy is needed. Authentic outcome and cleanup follow below.
Authentic result
Correction 467fb65b3 was published through ordinary merge e7f705046 after
the focused checks and locked full build passed. One fresh registered original
(native PID 2482839) used the genuine COMCTL32/COMMCTRL pair, unchanged registered
92-byte Settings100 and 64-path/62-manifest asset closure. Genuine COMCTL32
DllMain returned EAX 1. The actual main and Worker source/module readback audit
matches 232 reads across thread IDs 0, 1 and 2. The full genuine warning frame
was personally reviewed before the finite 300 ms Enter. Native exact-invocation
consumption and independent physical UP both succeeded.
The repaired retirement behavior is confirmed in the authentic application:
afterRelease and postProgress report focus NULL, rather than the destroyed IDOK.
Main Riot Engine HWND 0x10004 is live, generation 4, process 1000/thread 1.
All 183 retained pending responses before UNKNOWN contain exactly
focus.live, focus.generation, focus.threadIdentity, focus.liveIdentity,
focus.thread, focus.mainTopLevel. Ordinary USER owner, actor and process
checks do not fail. No guarded presentation becomes ready by the unchanged
3000 ms operation deadline; the scope closes UNKNOWN with no later endpoint.
This is the remaining concrete blocker. At the two retained snapshots the
renderer calls the main window visible while its native style is still
0x00ca0000 (WS_VISIBLE absent). Logs record ShowWindow(5), followed later by
ShowWindow(1). This is consistent with incomplete show/initialization work,
but does not prove the active callback, its cause, or an activation defect.
The authored native activation path passes. No early focus assignment, relaxed
guard, fabricated live focus, extended operation clock or game-specific fix was
introduced. A second original was not spent without another demonstrated,
tested and published correction. Neither predecessor scope was reopened.
Post-warning whole frame, menu, level, visible movement/attack and ordinary Quit remain unproved. Local cancellation completed actual native and independent watchdog waits with both EOFs and no leaked processes; cancellation exit 0 is not ordinary Quit. Native 900 s/watchdog 920 s deadlines and the 180 s Quit reserve were unchanged. Original minimum free space was 1,899,126,784 bytes; no RAM gate was used. Before launch a temporary shared disk shortage was reported and the launch held until space recovered. No box was created: ID null, state NOT_CREATED, expiry null, stop receipt null.