Dungeon Keeper bounded native motion, 7 November 2026

The native Pointer Lock path passed locally through the production renderer and guest input consumer. The single fresh original attempt stopped at its first blocker: the ordinary trusted capture click did not acquire Pointer Lock. The motion command refused before sending any movement. I reviewed the original intro and Main Menu only; no Eversmile dungeon, digging, building, action consequence, FPS, audio or parity was demonstrated.

Evidence and the retained browser adaptations are in scratch/runs/20261107-dungeon-keeper-continuous-native/. Predecessor caae9bb35 was published and released before this assignment started. Its published note and the Gameplay route/relative-cursor sections of dungeon-keeper-demo.md informed this test; its accepted PNG/capture fixture matrices were not repeated.

tools/dungeon-keeper-native-motion.cjs adds bounded ordinary CDP movement behind the existing Canvas2D chain's immutable reviewed-pixel admission. Commands allow 1–32 nonzero integer steps, each axis at most eight counts, intervals of 20–100 ms, and nominal movement duration at most three seconds. A five-second command deadline prevents further dispatch after slow diagnostics. Every step checks live process/window/document/module-object continuity, released buttons, the owning canvas's Pointer Lock and unchanged geometry; physical coordinates must remain inside the canvas. Full executing-module hashes are checked during admission. A failed/UNKNOWN dispatch closes the chain and is never replayed. There are no programmatic renderer movement calls or guest-memory writes.

The retained browser tracks physical CDP coordinates across commands: locked DOM clientX/Y stay frozen and cannot be the next command's origin. Locked clicks preserve that physical point and use the production input bridge's virtual guest cursor. The run-only transport now verifies the exact command number in both the returned and retained ACK, as well as its ID. Existing timestamp-before-readback freshness, 22 identity comparisons, natural publication, independent UP scheduling and sticky terminal behavior remain in place.

Local qualification used the normal production index route with the installed dx_ddex4 sample, Win98Renderer, Canvas2D and an actual guest Worker. Only this local sample's app profile was set to relativeMouse: true to exercise Pointer Lock. The original registry/configuration was unchanged. A passive wrapper of the real renderer.takeInput recorded the events returned to host check_input; it called the original exactly once and returned the same event unchanged. This is evidence of the guest input consumer, not a claim about game simulation or DirectInput (Dungeon Keeper's documented path uses WM_MOUSEMOVE).

I personally reviewed the fresh immutable sample captures: red sprites on a blue checkerboard with the F12 caption. In qualification/, two consecutive commands delivered eight trusted locked DOM mousemoves, each (2,0), with no spurious jump between commands. The actual consumer returned all eight WM_MOUSEMOVE events, HWND 65537, with coordinates 322, 324, …, 336 at Y=240; the queue drained. The preliminary local/ pass also delivered four such moves. Both browsers exited normally with code zero and removed their profiles.

Focused actual-chain controls in qualification/native-controls.json passed: stale and wrong-window-owner admission had zero effects; wrong numbered ACK and UNKNOWN movement lanes became sticky and rejected a further command. An actual mouse pulse armed UP before DOWN and attempted UP after 60 ms; UP completed before the deliberately delayed DOWN acknowledgement. These are new native motion/transport controls, not the accepted capture fixture matrix. test/test-dungeon-keeper-native-motion.js additionally checks invalid bounds, physical-coordinate continuity across commands, and terminal mid-command lock loss with no remaining sends. It and the required locked build passed. No emulator/renderer implementation was changed or defect asserted.

One UUID Idempotency-Key POST created bx_mp72p7pt. The isolated deployment retained the complete 644-pin predecessor PNG closure and added the motion helper, for 645 pins. All pins passed SHA/length verification before launch and after shutdown; stage.cjs --verify resolved the complete pngjs closure inside the deployed root. All 168 original media/support pins stayed unchanged. The registered dungeon_keeper_demo route, original media and configuration were preserved. Source identity is Git SHA 0029a47340d527ff1d65c8470ef16bb0781cd8ac plus the retained relevant binary patches, including new files. Existing media is referenced by path/hash, not archived. Small helper recovery patches are retained; no source/media archive was created. Page and guest executing module SHA-256 was e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467.

Fresh reviewed original captures are review/startup.png (Bullfrog vortex), review/menu.png (Main Menu, imp near Multiplayer), and review/locked-menu.png (still Main Menu; the acquisition click moved the software cursor to the lower-right edge). The last filename names the intended checkpoint; it does not establish that Pointer Lock succeeded. The owner-checked read at 14:43:04.664Z, recovered as passive-lock.json, reports software cursor (639,479), last native position (320,240), relative mode 1, no held buttons, an empty input queue, two active guest threads, and pointerLockElement: null.

Command 0's Escape and command 1's acquisition click have exact numbered ACKs and balanced trusted DOWN/UP events. The mouse UP was independently armed before DOWN and attempted exactly 100 ms later. Command 2 requested four two-count upward movements against the fresh reviewed menu. At 14:43:34.570Z the helper failed own released Pointer Lock required, before the first mouseMoved dispatch. The final native log contains only five events: Escape DOWN/UP and the acquisition move/DOWN/UP. There was no motion replay, substitute handler call, guest-memory write, configuration change or second original runtime. Browser/transport failure is terminal. The exact reason why the original did not acquire lock was not instrumented. Static code shows its unchanged registry uses automatic detection, unlike the explicit local sample profile; that difference is context, not proof of a browser or emulator defect.

The immutable 25-minute cap was fixed from allocation at 14:41:26.256Z to 15:06:26.256Z. Chrome exited normally at 14:43:35.165Z, code zero. Recovery verified 61 artifact paths / 2,937,384 bytes by SHA/length before normal stop. Custody verifies both owned PIDs absent, no matching live runtime, server/debug ports closed, profile/writer absent and held input empty. Delegated normal stop completed 14:44:19.214Z, forced: false; final state is ARCHIVED, current expiry none, original running expiry 15:21:26.108Z. No old boat was touched and no lifecycle ownership remains.

The local one-second monitor recorded minimum free disk 2,593,013,760 bytes; the remote browser's 2 GB guard recorded no floor event. Monitoring has sampling gaps, so this is not a continuous disk guarantee. Native qualification, unit tests, locked build, pin/media checks, retrieval/custody, owned evidence checks and documentation whitespace passed. There is no candidate manifest entry for this original, so evidence stays in these notes rather than attaching menu screenshots to an unrelated dashboard candidate or a gameplay badge. The global evidence checker still reports the preexisting unrelated Alice containment error and historical missing artifacts; none belongs to this run. The local recovery helper was amended to persist individual cleanup predicates before asserting them; this did not change the deployed runtime pins.

Next work is to qualify bounded trusted continuous motion when the unchanged original route does not acquire Pointer Lock, or establish the native capture prerequisite without changing its configuration. This assignment stops at that exact first blocker; another original runtime requires separate authorization.