Drakan post-warning startup readiness — 2026-11-18

Evidence: scratch/runs/20261118-drakan-post-warning-startup-readiness/. Every predecessor UNKNOWN scope remains closed. No predecessor is resumed.

The input helper coupled completed warning input to the next focused capture. A real ordinary authored PE reproduces this: NULL-owner MessageBox, native Enter consumption and independent physical UP, 4500 ms ordinary GetTickCount-based initialization, hidden main creation, ShowWindow(5), a 700 ms WM_SHOWWINDOW callback, DefWindowProc, and the ordinary message loop. The old workflow expires its unchanged 3000 ms operation clock although DOWN/UP succeeded. The separate workflow completes that same input on time, then observes passive startup under a separately predeclared finite bound. The readiness deadline is sealed at input completion, never restarted by the observer. Only this reviewed pre-main warning Enter can request separation; other inputs retain their original behavior.

The passive phase calls existing native-startup-readiness and read-only actor diagnostics. It allocates no capture/input lease, never pumps or wakes the guest, never assigns focus/visibility, and never changes guest state or clocks. Every reply, failure list and selected actor/stack observation is retained before the next request. Process/screen/physical-UP/actor identities remain checked. Once ready, capture has its own unchanged 3000 ms operation and existing identity and freshness guards. A timeout closes the task's scope with no later endpoint.

Native oldfail/newpass controls run the same authored PE/module. Nine identity mutations are refused. The normal browser shell also runs the authored program in a real Chromium Worker, with actual DOM Enter DOWN/UP. The Worker acquires live main focus after ordinary delayed initialization/show, while the idle page instance retains NULL. An initial reduced browser harness omitted ordinary host scheduling and timed out; it is retained as a failed fixture, not a runtime finding. The passing control uses the full normal production shell and scheduler.

Native evidence records the SWND continuation cookie and WM_SHOWWINDOW frame before activation, with hidden native style and correct thread 1. This separates unfinished callbacks from wrong-instance focus observation. No runtime defect is inferred from renderer/native visibility disagreement alone. This change is a tested preparation correction, not a claim that Drakan's remaining runtime cause has already been established. Authentic outcome follows after publication and the one authorized fresh registered original.

Published preparation correction: 2b8ba1b67. Focused native controls, the full-shell Chromium Worker control, identity negatives, existing native focus, guarded input and authored activation controls pass. The locked full build passes. No shared runtime or foreign-owned input/shell file was edited.

The fresh registered original confirms the causal preparation defect. Native Enter consumption and independent UP completed in 459 ms, within the original 3000 ms operation. Main focus became ready 12.356 seconds after input completion, after 114 passive replies. Every pending reply had exactly the six previously observed focus failures, with the correct process/thread actor. Ordinary native ShowWindow(1) then produced live Riot Engine 0x10004, thread 1, generation 4, style 0x10ca0000 (WS_VISIBLE), and real main focus. The observed delay is legitimate initialization; renderer visibility did not mean activation had completed. Read-only actual Worker stacks/EIPs show initialization progress. The authored delayed WM_SHOWWINDOW control separately records the SWND callback continuation before ordinary activation. There is no evidence requiring another runtime activation fix.

Whole-frame review found a black first post-warning frame, then the blue dragon loading medallion. No menu, level, movement, attack or ordinary Quit was proved. At 04:49:47 UTC the independent watchdog stopped the original because actual free disk fell to 1,484,599,296 bytes, below the required 1.5 GB floor. Native SIGTERM performed orderly shutdown (exit 0); native and watchdog actual waits both retain stdout/stderr EOF, and containment reports no leaks. Native final statistics show 62,922 batches and 123 presents; rendering activity alone does not establish visible gameplay. This resource stop is the concrete unresolved blocker. The coordinator then explicitly authorized the one extra original after its lossless disk recovery restored headroom. All predecessor scopes remained closed.

A subsequent passive helper attempt was refused locally by opened() because the native wait already existed; no endpoint was sent. No closed runtime was queried or replayed. The source/module actual-read audit, unchanged media and Settings checks, selected whole PNGs, waits and recovery records are retained in the run directory. result.json is written last. No test box was created (ID null, state NOT_CREATED, expiry and stop receipt null).

The affected extra original verified over 2.3 GB free and used the unchanged published preparation/module. Its first capture was mistakenly invoked before the detached owner had published native-pid.json. The local identity guard refused ENOENT and closed the helper before sending any endpoint request. This is a launch-ordering preparation error, not a native/game finding. The native process was subsequently stopped by identity-checked SIGTERM; both actual waits and EOFs are retained, exit 0 and no leaks. No capture or input reached this extra original. The coordinator's later 04:54:07 same-task clarification explicitly authorized one fresh replacement after a tested correction, without reopening either closed scope.

tools/drakan-launch-readiness.cjs now bounds filesystem-only launch admission: the detached owner, live native identity and armed watchdog receipts must exist before capture can be invoked. Early terminal receipts or expiration refuse admission without endpoints. The authored compiled native PE test demonstrates the old immediate read failing before delayed receipt publication, followed by new verified admission and actual warning execution. Real supervised early exit and timeout controls refuse admission; every scope retains both actual waits and EOFs. This does not change any input/capture/runtime deadline. The replacement's outcome will follow tested publication. The earlier error does not invalidate the first original's measured input/startup separation.

The receipt-order correction was tested, passed the locked full build and was published as 5eeed9ab8 before the explicitly authorized fresh replacement. Both its launch recipe and operation entry point wait for verified native receipts before capture. This succeeds with native PID 2540351. Warning Enter and independent UP complete in 437 ms; 103 passive samples establish ordinary main focus 10.783 seconds later. All 233 actual source/module reads match the sealed inventory, on main and Worker thread IDs 0, 1 and 2. The module remains a51912dd4922e26a02689ede7e12cab8d8400b1649ab643e4918ad61f4e8dc0a.

Personally reviewed complete frames establish the Psygnosis logo, Surreal intro, Drakan Demo 1.0 snowy title and finally the authentic main menu at batch 96,746. Its central New Game, Load, Save, Multiplayer, Options, Credits and Quit controls are visible. This establishes the requested authentic post-warning menu through ordinary initialization/show/focus, without a guest/runtime patch, forced visibility/focus, forced wake/poll, pause or wider clocks.

One finite 300 ms Enter was then attempted from that reviewed menu to select New Game. The guarded dispatch timed out; the independently prearmed release also timed out and records UNKNOWN. Local native logs show WM_KEYDOWN and WM_KEYUP for Enter reaching focused 0x10004, which does not prove either native consumption or physical-UP acknowledgment. No later endpoint, capture, input retry or ordinary Quit was requested. The runtime scope is CLOSED_UNKNOWN. Identity-checked SIGTERM cleanup retains actual native/watchdog waits, both EOFs, exit 0 and no leaks. Final native statistics are 106,144 batches, 502 presents and 479.943 seconds. Level, movement, attack and ordinary Quit remain unproved. The concrete remaining blocker is guarded input/release acknowledgment at the authentic menu; its cause is not established by those local delivery logs. No claim is made that the menu Enter selected New Game, or that the renderer/native activation contract needs another fix. All three original lifetimes and every predecessor scope remain closed.

During final verification all seven selected PNGs disappeared from disk, also reported independently by other active workers. The deletion's actor is unknown. Each image was recovered byte-for-byte from this task's persisted image-tool response in its own Codex session log, matching every previously recorded SHA256 and size. recover-reviewed-images.cjs and image-recovery.json retain exact record-line/hash recovery provenance. No runtime, endpoint or capture was used for recovery. The existing unique runtime module remained available and matched its pinned hash. Selected PNGs are retained again; recovery does not invent or re-render evidence.

The two large original native stdout logs also disappeared during finalization. They are explicitly missing; complete logs cannot be reconstructed. Exact previously returned tool-output excerpts preserve ordinary ShowWindow, menu DOWN/UP delivery and final statistics (native-log-excerpts-final.json), with source record hashes. Full passive readiness/actor JSONL, actual readbacks, guard receipts and native/watchdog wait/EOF receipts remain. No full-log retention claim or runtime replay is made.