Dungeon Keeper corrected native original, 7 November 2026
The corrected retained pointer driver passes on the actual original: small trusted motion reaches the normal consumer without the former fractional anchor mismatch. The deepest personally reviewed checkpoint remains Main Menu, with credits also observed and dismissed. Start New Game, Eversmile, digging, building and gameplay performance are unproven.
The terminal demonstrated blocker is a live automatic capture transition: after 22 small vertical steps, guest recentering changed capture intent to true while Pointer Lock remained absent. The unchanged helper correctly refused step 23 and closed the lane. No replay, anchor reset, forced profile, private input handler, guest write, clock/config change or production repair followed.
Evidence: scratch/runs/20261107-dungeon-keeper-corrected-original/.
Predecessors: anchor conversion 9a0975685 and capture gesture ee103ca06.
Preflight and actual original
The retained 20261107-dungeon-keeper-unlocked-native/browser.cjs and
tools/dungeon-keeper-driver-pointer.cjs were used. All 647 staging paths
had lengths and SHA-256 hashes recomputed, including the new pointer
dependency. They matched locally before allocation and remotely after the run.
Source SHA was 9a0975685e487833bc701d69418e2cbd7ce27127. The executing
guest/page module, recovered through CDP, was
e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467
(2,182,493 bytes). Exact browser versions are in the recovered version record.
One local production dx_ddex4 preflight exercised the corrected driver,
full-identity observer and numbered transport together. I reviewed the three
red sprites/checkerboard captures. A neutral gesture submitted CSS Y246.8,
recorded trusted MouseEvent Y246, then four trusted (2,0) steps reached the
ordinary consumer at X321,323,325,327/Y52. Exact ACKs 0–2, independently armed
UP, released inputs and browser/server/port/profile cleanup passed. This is
fixture input evidence only. Accepted matrices were not repeated.
Two local preparation mistakes are retained: an initial background launch
ended before browser creation, leaving only its execution-window file; the
local observer initially used its default Dungeon Keeper app check and
refused before installation. Its local-only adapter then selected the actual
registered dx_ddex4 owner. The remote observer retained the original app check.
Observer return/throw/once/restoration controls and syntax checks passed.
A separate bounded watchdog process was delegated before the sole allocation.
One UUID Idempotency-Key POST created bx_gq2jk9k6; no old box was resumed,
and no POST was repeated. The original registered dungeon_keeper_demo
route used production automatic capture selection, with no explicit profile.
The observer reread all 22 owner/module/window/epoch identity fields before
and after observations, including executing module bytes. Its later cursor
diagnostic only added reads of documented guest variables.
Reviewed pixels and consumed input
I personally reviewed the immutable owner captures in review/:
startup.png: red rotating intro. Ordinary Escape DOWN/UP, ACK 0, visibly led tomenu.png, Main Menu.after-gesture.pngandmotion-qualified.png: dim Main Menu, following the neutral title gesture and four motion qualification steps.start-hover.png: credits, despite its intended navigation label.after-start-new.pngandnew-game-result.png: Main Menu, with the imp cursor still near the same central position. Neither proves Start New Game.cursor-baseline.png: credits.menu-for-continuous.pngis a dim credits transition after Escape;settled-menu.pngis the restored Main Menu.
There is an explicit review error in this run: command 4 was submitted with
an incorrect Main Menu description after start-hover.png actually showed
credits. That click returned to Main Menu, rather than proving Start New Game.
The next command's review records the correction; the original records were
not rewritten. The initial menu review also incorrectly enumerated a Credits
row; the next motion review corrects it. These mistakes limit the navigation
claim. No Eversmile or gameplay screenshot exists.
The neutral ordinary gesture at normalized (0.5,0.11) had automatic false
intent, no lock and no requestPointerLock call. Trusted MOVE/DOWN/UP reached
HWND 98305 at (320,52), with UP independently armed before DOWN. Command 2
sent four (2,0) steps: native X354,356,358,360/Y246, ordinary consumer
X321,323,325,327/Y52. Command 3 delivered ten further small steps toward the
Start New row. These establish useful coordinate delivery, but do not prove
software cursor steering or successful menu selection.
passive-cursor-before.json, under the full identity check, reads software
cursor 0x761378/0x76137c = (320,240), last native position
0x762b70/0x762b74 = (320,127), and private relative flag 0x51b674 = 1.
Renderer position was (320,127), queue empty and buttons zero. Automatic
selection was false. The private flag was not substituted for capture intent.
This demonstrates delivery/last-position versus software-cursor divergence at
that checkpoint. It does not establish a semaphore or emulator root cause.
After the reviewed settled Main Menu, command 9 requested 24 (0,-4) steps
at 100 ms intervals, from the unchanged native anchor (352,321). The helper
sent and acknowledged 22 events, CSS Y317 through 233; ordinary consumer
Y123 through 39, X320. The last two consumed events had cursor-warp serials
1 and 2. At 15:34:25.730Z the prospective guest (320,35) selector returned
true, with a recorded recenter run 3 to (320,240) and a two-second recenter
window. Lock was absent. The helper refused before sending CSS Y229/step 23:
motion capture intent or guest picture changed. No actual step 23 or24
exists. This was not an anchor conversion failure, geometry escape or
recorded Pointer Lock rejection.
Exact successful numbered ACKs were 0–8. Command 9 has no ACK; first-failure and sticky transport-closed records remain. The browser closed on that first fault, so a subsequent observer attempt found no live endpoint and the subsequent retrieve found the already retained settled capture. Those errors are preserved; no lane was reopened. Final software cursor after this partial command is unknown, and no post-motion screenshot was obtained.
Executable next diagnostic and limitations
Run the new offline diagnostic, which opens no browser and sends no input:
node scratch/runs/20261107-dungeon-keeper-corrected-original/next-transition-diagnostic.cjs
It feeds the recorded production selector transition through the unchanged
helper, reproduces the 22-event boundary and asserts that step 23 is unsent.
This check passed. A separately authorized fresh original should trace an
ordinary trusted capture gesture during the measured recenter window, with
fresh personally reviewed owner pixels and the existing full identity gates.
Observe actual request/change/error results; require lock whenever intent is
true. Do not resume this lane or infer acquisition from the private flag.
Only after actual software-cursor steering should a successor select Start
New Game and attempt Eversmile/dig/build. The cursor commit path documented
in dungeon-keeper-demo.md remains a further diagnostic target if acquisition
succeeds but the software cursor still fails to move.
No demonstrated production-code cause justified a fix here. No production code changed, no new build is claimed, and the predecessor's locked build remains the tested module. The global evidence checker still reports unrelated Alice/Tunnel-Twist containment errors and historical missing artifacts; this run has no candidate-manifest registration. Own custody/hash assertions pass.
Source recovery and terminal custody
source.json, runtime-pins.json, relevant tracked/new binary patches and
owned-inputs.patch record recovery. All 168 existing media paths have length
and SHA references; source/media archives were not created. The 136 unique
retrieved artifacts total 5,092,977 bytes, each verified against remote length
and SHA. All 647 remote runtime pins remained unchanged.
Chrome exited zero. Browser/server PIDs 5913/5937 are absent; ports 40497/46157 are closed; profile, writer lock and matching runtime processes are absent. All native inputs are released; final buttons are zero and queue empty. Closing the owned pages destroyed their observation wrappers. Local preflight browser/server/ports/profiles are closed, and the owned board watcher and delegated lifecycle watchdog have ended. Sampled free /home stayed above 2 GB.
The box's initial archiveAfter expiry was 16:10:15.820Z. The browser's
immutable cap was 24 minutes after allocation, inside the watchdog's 25-minute
allocation cap. After verified retrieval, one normal stop
stop_cf0b04c3298d9faf completed 15:35:20.233Z, forced=false, error null.
Box bx_gq2jk9k6 is archived, stoppedAt 15:35:20.205Z. Its current
archiveAfter is null; retained-data expiry is unknown. Shutdown was explicitly
delegated by this task; no assigned running box remains.