Dungeon Keeper measured recenter acquisition, 7 November 2026

The fresh original reached personally reviewed Main Menu. Acquisition is blocked by a demonstrated timing mismatch in this guarded test path: its guest warps were 533.99, 511.22 and 520.675 ms apart, whereas production automatic selection requires matching warps strictly less than 500 ms apart. Each warp therefore reset to run 1. Automatic intent stayed false; the command correctly sent no acquisition DOWN and closed its lane. Pointer Lock, Start New Game, Eversmile and a visible dig/build action remain unproven. This is a harness acquisition blocker, not a demonstrated emulator cursor defect or a Pointer Lock rejection.

Evidence: scratch/runs/20261107-dungeon-keeper-recenter-acquisition/. Predecessors read: corrected original 4a95f08d4, capture gesture ee103ca06, anchor conversion 9a0975685, and the cursor commit path in dungeon-keeper-demo.md. No accepted matrix was repeated.

Exact path and local closure

The retained native driver, observer and numbered transport were derived into small run-specific inputs. The driver still uses the corrected trusted MouseEvent anchor conversion and unchanged tools/dungeon-keeper-native-motion.cjs. The new acquire.cjs executes at most four one-step calls through that helper, observes the actual production selector and recenter state after each successful call, and stops sending motion as soon as intent is true. Only a live run >=3 with an unexpired window allows an ordinary trusted mouse DOWN/UP. It never directly requests capture. UP uses the retained pulse, armed independently before DOWN. Refusals and UNKNOWN outcomes close the outer driver/transport; no lane is reopened.

All existing owner/module/window/epoch/freshness gates, exact native anchor, automatic profile, and identity comparisons remain. Observer installation and observations compare all 22 identity fields before and after, including executing module bytes. Acquisition adds another full live identity read between successive helper admissions. No forced profile/capture, anchor reset, guest write, private input handler, clock/config change or UNKNOWN command replay occurred.

One actual local production dx_ddex4 fixture qualified the exact derived driver and observer with ordinary trusted native input. I personally viewed its startup, gesture and motion PNGs: three red sprites over a blue checkerboard. The neutral gesture and four (2,0) steps reached the normal consumer at guest X321,323,325,327/Y52, receiving exact numbered ACKs 0 and 1. The no-intent acquisition control then sent three guarded one-step motions, with step ACK gaps 466 and 468 ms, but no DOWN. Its expected missing-window failure closed command 2. Chrome exited zero, inputs released, queue empty, browser/server PIDs absent, ports closed and profile/writer lock removed. An initial background shell launch produced no fixture/browser evidence; the subsequent retained foreground execution supplied this local closure.

test-acquire.cjs separately runs the production selector with the retained motion helper and pulse in an offline model. Three matching warps stop the fourth move and admit ordinary DOWN/UP. Absent/expired intent, a helper refusal and an unknown dispatch all prevent DOWN. This positive model supplies no actual browser Pointer Lock evidence. The actual local fixture never armed capture intent. Its sub-500 ms timings did not predict the original's timing.

One original and reviewed consequences

A separate watchdog process was explicitly delegated before allocation, with a 25-minute cap. One persisted UUID Idempotency-Key POST created bx_xxzdzgs4. No old box was resumed, no allocation POST was repeated, and no other original ran. The registered dungeon_keeper_demo route retained automatic capture selection. All 650 runtime pins were hashed/length checked locally before allocation and remotely after execution.

I personally reviewed the immutable owner captures before each input command:

The neutral-gesture review text mistakenly called Multiplayer “1 Player”. The acquisition review explicitly corrects it; the original record is preserved. No menu entry or gameplay success was inferred from that text.

The full-identity passive reads retain an actual software-cursor consequence: before Escape/gesture, software cursor (320,240) and last native (0,0); after the neutral gesture, software cursor (639,292) and last native (320,52), relative-mode flag 1. Thus this original did not reproduce a universally immobile software cursor. The result is consistent with the documented delta-from-last-position commit and clamping, but no API/commit trace establishes that causal interpretation here. Divergence alone is not an emulator defect. Useful continuous steering remains unproven.

Acquisition command 2 completed four successful one-step helper calls. Native CSS Y values were 242,238,234,230 at X352; consumed guest positions were (320,48), (320,44), (320,40), (320,36), with warp serials 0,1,2,3. The four post-step observations recorded center warps (320,240) at performance timestamps 54732.690,55266.680,55777.900,56298.575 ms, each run 1. Every selector result was false; no recenter deadline was armed. Repeated full admission plus full observation between one-step calls consumed the available 500 ms interval. The recorded predicate explains the reset without assuming a browser capture failure or semaphore starvation.

At 15:46:16.989Z the bounded command failed measured automatic recenter window absent; no DOWN. Exact successful outer ACKs are 0 and 1; command 2 has no ACK and became sticky UNKNOWN. All four move ACKs are retained inside its acquisition receipt. No additional motion, gesture or input was attempted. The acquisition DOWN was never sent. The installed observer recorded no Pointer Lock request/change/error before the acquisition command; its final rows after that command were not recovered before browser closure. Actual lock is absent in every acquisition observation and terminal input state. No request/rejection is claimed from missing rows. Final software cursor after these four moves is unknown, because failure closed the browser before another full cursor observation/capture. Queue consumption is not substituted for that missing consequence. There is no post-acquisition screenshot and no qualified locked steering.

Executable next diagnostic

node scratch/runs/20261107-dungeon-keeper-recenter-acquisition/next-timing-diagnostic.cjs

This offline diagnostic passed. It checks the production threshold predicate, reconstructs the four run-1 resets from the immutable observations, checks consumed points, no DOWN, exact failure number, missing ACK 2 and sticky transport closure. It opens no browser, changes no clock and replays no native command. Its JSON output records the next local qualification: batch passive selector/warp observations with a bounded retained-helper group at the edge, preserving full owner/module reads before and after and the existing per-step progress gates. Demonstrate all actual guest warp gaps below 500 ms, stop without an extra move when intent changes, then observe an ordinary trusted DOWN during true intent. That local path must qualify before another original. This archived box and command 2 cannot be reused.

No concrete emulator cause was demonstrated, so there is no production fix, fail-before/pass-after emulator regression or new build claim. The mandatory locked-build requirement would apply to a justified code repair. The prior anchor correction's built module remains the executing module here.

Recovery, checks and lifecycle closure

Source SHA: ecf85c4d92346f277f02d7142be3d3ef0b0f394a. Executing page/guest WASM, recovered through CDP: e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467, 2,182,493 bytes. Browser: Chrome 151.0.7922.108 / V8 15.1.206.16. source.json, runtime-pins.json, relevant tracked/new binary patches and owned-inputs.patch supply recovery. All 168 existing media paths have length/SHA references in mediarefs.json; no source/media archive was made.

Focused acquisition controls, syntax, timing reproduction, exact ACK/pulse assertions, own evidence custody/hashes and whitespace checks passed. The global evidence checker still reports unrelated Alice/Tunnel-Twist containment errors and historical missing files. This app has no candidate-manifest run registration; evidence remains in this investigation note.

All 64 retrieved unique artifacts, 3,015,475 bytes, match remote length and SHA. Chrome exited zero; browser/server PIDs 6342/6349 are absent; ports 40415/39623 closed; profile/writer lock absent; no matching runtime processes, held input or queued input remain. Local fixture resources, own board watcher and delegated lifecycle process ended. Closing the pages destroyed observers. Sampled free /home remained above 2 GB.

Initial box expiry was 16:24:30.419Z; browser deadline was 24 minutes after allocation, inside the delegated 25-minute cap. After verified retrieval, normal stop stop_b1829bbee455cd99 completed 15:46:34.372Z, forced=false, error null. Box bx_xxzdzgs4 is archived, stoppedAt 15:46:34.346Z. Current archiveAfter is null; retained-data expiry is unknown. Shutdown was explicitly delegated by this task; no assigned running box remains. No MW3 or graphics-worker path was modified.