Drakan native beta-warning bootstrap and current gameplay, 2026-11-08
The pregraphics warning is now locally qualified through an ordinary native Enter, followed by real production D3DIM drawing and a one-way handoff to strict renderer binding. The original software arm reached the rainy level through native warning OK → relative-input New Game → Normal → story Enter. Outgoing attack success remains unproved; incoming enemy damage is not counted as player attack.
Evidence: scratch/runs/20261108-drakan-native-warning-gameplay/ (ignored). Predecessors: current browser action, demo notes. Only this note is committed; there is no emulator code change.
Qualified local path
The generated 25,088-byte x86 guest executes production MessageBoxA/GDI, then ordinary DirectDraw7/Direct3D7 COM calls to draw a red TL triangle and present it. It uses no private dialog command, forced return, API injection, warning bypass or Q3 GL substitute. generate-fixture.cjs and its exact input patch are retained.
The bounded dialog phase checks the sole app/process, exact CPU Worker target/endpoint/slot, recovered actual module, native modal identity/bounds, native OK child/class/bounds/focus and stable page object. Renderer absence is expected only here. A reviewed Enter100ms has an independent release timer and exact numbered ACK. After release, actual renderer creation must pass unchanged strict identity checks plus two advancing, exact-owner D3DIM draw/completion samples. This permanently closes dialog admission; postbinding replacements remain fatal.
local-software-qualified/ and local-webgl-final/ pass real warning input, visible production D3DIM triangle, strict posthandoff native input and all controls. Dialog wrong-owner/module/stale/window/replacement/UNKNOWN observations produce zero dispatch and sticky closure. Graphics zero-draw/missing-ready/wrong-owner/module/replacement/stale/UNKNOWN controls produce zero dispatch; graphics stale alone permits only a fresh capture. All other faults remain closed. local-pointer/ additionally proves trusted native relative movement and independently released mouse down/up, with final zero buttons. These boundary controls alter observations, never guest state.
The first local WebGL readiness attempt failed because the adapter exposed neither .gl nor .stats. The corrected passive observer retains the actual D3DIM GPU created by the production adapter and verifies endpoint plus instance, live nonlost context and advancing actual GPU snapshot counters. The same production calls execute exactly once unchanged. A rerun passes both backends. This is a run-helper correction, not an emulator repair. Earlier setup failures remain recorded. An early local software drain check ran before the driver’s eight-second shutdown wait; subsequent full drain checks confirm completion, and original backend arms are strictly serialized.
Original fixture and ownership
Tested source is 9fd4b4a5cf14d8afd8665df0a8e2b34f79dd66ca; relevant-source.patch, relevant untracked source patches, owned run-input patches and RECOVERY.md provide recovery without archiving source. The unchanged registered 64-file fixture/media and startupRegistry are pinned by exact existing paths/SHA256 in media-preflight.json and runtime-pins.json (681 runtime files). Historical per-file media digests were unavailable. Actual CPU and render Workers both recover module SHA256 f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121, 2,182,500 bytes.
One UUID-idempotent fresh box, bx_y3ptseub, was allocated at 04:02:09Z with an independent observable supervisor/watchdog prearmed before allocation, deadline 04:27:09Z. Browser, work, ACK and cleanup deadlines remain bounded; no RAM gate or config/clock/state forcing was used. Disk remains above five billion bytes. No old box was resumed or deleted.
Software warning ownership is process1000, CPU endpoint1/slot0, modal65537 at (40,40,334,124), native focused OK child65539 at (168,127,72,24), with no renderer. Ordinary Enter creates the actual software D3DIM renderer endpoint1. CPU target 9733B64E18D99500F6039CC4E636333B and render target 61D6083222430A8E5F9F489AEA5F3518 persist through the arm. Complete pre/post identities are in capture/binding/handoff records.
Reviewed software action
A trusted click acquires the production-requested pointer lock; prelock movement is not claimed as game movement. Native relative (-6,-75) then visibly hovers New Game with its tooltip. New Game click, relative (-2,+69) and reviewed Normal click reach the story/loading screen; ordinary Enter reaches the rainy outdoor level with Arokh, river, ridge and HUD.
One Normal click was refused stale with zero effects and an exact ACK, followed only by fresh capture before input. Two human review prose errors are preserved with correction records: warning bounds were mistyped in prose although machine-bound bounds were correct; a New Game pressed frame was prematurely described as difficulty, which was actually confirmed in the next capture. No consumed review was rewritten or input replayed.
ArrowUp2000ms and native relative horizontal -100 have exact ACKs and independent release. Ridge/tree/bridge positions change across story-enter, forward, turn-left and attack; the substantial leftward heading change appears after the mouse move. This supports current terrain-relative heading change, but not sustained flight or altitude change. Enemy motion and rain alone are not movement proof.
Primary left mouse1500ms and secondary A2000ms were delivered against visible nearby enemies. Neither reviewed postinput nor secondary held frame shows an outgoing breath/projectile and a target consequence. Enemies visibly strike Arokh and the health orb empties; these are incoming damage. Do not promote this run to attack acceptance. Software ends with 20 commands and 20 exact ACKs, no UNKNOWN; Chrome exits 0, driver/ports/profile/writer drain, and all681 pins remain unchanged before the next arm.
Reviewed WebGL action and remaining blocker
After complete software drain, the same box ran WebGL serially. A launch-template typo accidentally renamed the deadline environment key; the process failed before creating its output directory, Chrome or any guest. webgl-launch-correction.json preserves the cause; the corrected launch retained the original immutable deadline. This was not a guest UNKNOWN or input replay.
The actual CPU target is 087054066EB851AE5CBAFE9F3C73119B, render target 0BA786A58F478AD1FCF7E0E12E82DBDD, endpoint1, and live owning context token 153ad5db-7eb2-4152-825b-01d562a06a56:3; both modules match the pinned f965 digest. Renderer is ANGLE Intel UHD Graphics620/Mesa, WebGL2. The corrected exact-owner D3DIM projection proves advancing GPU drawCalls and completion, not merely a page-wide ready bit or unrelated GL activity.
Native warning Enter, bounded startup Enter activations, pointer lock, relative New Game hover, balanced New Game click, relative Normal hover and balanced Normal click reach story; Enter reaches the outdoor level. Captures sometimes precede visible application of a native event; passive settle captures confirm menu hover rather than inferring game behavior from host ACK. All input remains ordinary CDP.
Up Arrow2000ms produces no clear displacement in the reviewed immediate frame. Native horizontal -100 followed by passive settle produces a clear terrain-relative heading change: ridge, trees and bridge shift while health remains full (webgl/forward.png, turn.png, target-wait.png). Rain and dragon animation are excluded from that finding. No clear nearby target approaches during the bounded observation, so no blind WebGL attack is promoted. Software’s targeted primary/secondary attempts have no visible outgoing consequence before death. The precise remaining blocker is successful outgoing attack unproved; sustained flight/altitude and sustained forward movement also remain open. A bounded physical hold and producer/CPU counters do not by themselves prove that a game DirectInput poll consumed the button. No polling/hold ceiling was weakened.
WebGL ends with 19 commands/19 exact ACKs, no UNKNOWN; software 20/20. Both Chrome exits are0 and both terminal audits prove all driver/Chrome PIDs, ports, profiles and writer locks absent, with 681 unchanged pins. There was no original-arm overlap. FPS, audio, parity, full27 and full game acceptance remain separate.
Evidence and shutdown
remote-inventory-after-drain.json, retrieval-final-verification.json and retrieval-verified.json record postdrain retrieval and independent full byte SHA256/size verification before normal shutdown. All 459 references (404 unique payloads) were independently reread and SHA256/size verified before stop. Additional checks validate 108 PNG CRCs and decoding and 454 full HTTP200 payloads against staged pins. Box bx_y3ptseub archived at 04:23:43.103Z; completed normal stop receipt stop_0721d1a85dd5b797, forced=false. Final expiry/archiveAfter is null (initial expiry 04:27:09.223Z). Supervisor ends with no children and both supervised exits 0. See lifecycle records and result.json. The independent watchdog remains observable until normal stop confirms, then both children and supervisor exit. No public deployment, source archive, repeated media archive, box resume or deletion was performed.
Next: qualify evidence of actual DirectInput polling/attack timing locally, then use the retained native bootstrap for a new authorized targeted-attack attempt. Do not label level rendering, terrain heading or incoming damage as attack success.