Dungeon Keeper driver anchor conversion, 7 November 2026

The demonstrated fractional-anchor blocker is corrected in the retained native CDP driver. Its bookkeeping now records the trusted MouseEvent client coordinates, retaining the submitted CDP doubles separately as sentX/sentY. The continuous helper, production input bridge, automatic capture selection, identity checks, freshness, numbered ACKs and terminal UNKNOWN behavior are unchanged. This qualifies the local input path; Dungeon Keeper steering itself still requires a separately authorized original run.

Evidence: scratch/runs/20261107-dungeon-keeper-anchor-conversion/. No original, remote box, allocation or lifecycle operation occurred. Box ID/state/expiry are not applicable.

Measured conversion and change

The retained next-anchor-diagnostic.cjs from the capture-gesture investigation reproduced the pre-motion refusal with zero effects: driver (352,246.8) versus trusted MouseEvent (352,246). The equality gate was correct.

One small real Chromium HTML fixture (conversion.cjs) records both trusted PointerEvents and MouseEvents from ordinary CDP motion. Chrome 151.0.7922.108 / V8 15.1.206.16 converts submitted doubles to float positions; PointerEvent client coordinates preserve that float, whereas MouseEvent client coordinates floor it. Examples from conversion.json:

CDP input PointerEvent client MouseEvent client
352, 246.8 352, 246.8000030517578 352, 246
352.9, 246.1 352.8999938964844, 246.10000610351562 352, 246
351.999999, 246.999999 352, 247 352, 247
-0.9, -0.1 -0.8999999761581421, -0.10000000149011612 -1, -1

The near-integer case distinguishes floor(fround(value)) from flooring the original double. Deeper negative positions outside the viewport produced no document motion event; the fixture does not infer trusted coordinates there. The same fixture checks a canvas at CSS (32.5,194), CSS size 640x240, backing size 320x480. DOM coordinate conversion precedes canvas offset and scaling; guest mapping is a separate floor operation. This fixture was run during discovery and again with assertions against the correction. The original remote predecessor used Chrome 151.0.7922.71, not this local patch version.

tools/dungeon-keeper-driver-pointer.cjs implements only this measured bookkeeping. The retained scratch/runs/20261107-dungeon-keeper-unlocked-native/browser.cjs now invokes record(pointer,method,params) in its existing send path instead of assigning the doubles directly to pointer.x/y. Submitted event parameters are unchanged. This records each attempted event as before; an unknown dispatch still closes the lane. No anchor is read back and adopted, reset, or substituted after a mismatch. Locked physical continuity and unlocked exact equality remain in the unchanged helper. The conversion is scoped to this Chromium MouseEvent driver, not a general PointerEvent or guest-coordinate conversion.

Regression and actual production qualification

regression-before.cjs runs the new regression using the previous driver's double assignment. It exits 1 at the expected 246.8 !== 246 assertion (fail-before.log). The corrected regression passes. It covers positive and negative fractions, float precision at an integer boundary, unchanged submitted parameters, six repeated one-pixel steps across two commands, wrong-anchor zero effects with sticky closure, every one of the 22 identity equality fields, missing identity fields, and actual-lock versus automatic-unlocked capture admission. Existing native-motion tests also pass.

One local actual-production dx_ddex4 guest and its owning page Canvas2D renderer qualified the corrected driver and unchanged continuous helper. prepare-local.cjs derives the bounded local driver from the retained driver, using existing input/transport/readiness controls. It selects only this local registered fixture and preserves production automatic profile selection. I personally viewed local/startup.png, anchor.png and motion-one.png: three red sprites over a blue checkerboard. No Dungeon Keeper scene is claimed.

The canvas rect was (10,194,684,480) with 640x480 backing. The ordinary normalized (0.5,0.11) gesture submitted (352,246.8). Trusted MOVE/DOWN/UP reported (352,246); normal consumer events reached HWND 65537 at (320,52). Two commands then sent four (1,0) CSS steps apiece, with a fresh personally reviewed owner capture between commands. All eight trusted events had movementX=1, movementY=0, buttons zero, client X 353 through 360 and Y 246. Consumed MOVE X values were 320,321,322,323,324,325,326,327, Y 52. The first one-pixel CSS step remains within the same guest pixel because of scaling.

The automatic selection was false, no explicit profile was present and no lock was held. When capture intent is true the unchanged helper still requires lock; only automatic false intent qualifies unlocked motion. Actual chain admission reread the complete page/guest owner/module/window/document identity, including recovered executing module bytes. Negative controls used isolated lanes: stale scene and wrong window identity refused with zero effects; wrong numbered ACK and unknown motion closed admission without replay. The trusted independent UP was armed before DOWN and acknowledged at 15:20:56.926Z, before the deliberately delayed DOWN acknowledgement at 15:20:57.025Z. Exact successful command ACKs were numbered 0 through 3. All controls passed; no held retry or private input handler was used.

Verification commands:

node test/test-dungeon-keeper-driver-pointer.js
node test/test-dungeon-keeper-native-motion.js
node scratch/runs/20261107-dungeon-keeper-anchor-conversion/conversion.cjs
node scratch/runs/20261107-dungeon-keeper-anchor-conversion/validate.cjs
flock /tmp/wine-build.lock bash tools/build.sh

The locked build passed. The global evidence checker continues to report unrelated Alice/Tunnel-Twist containment errors and historical missing files; this investigation has no candidate-manifest registration and is documented here. Local assertions and custody checks passed.

Recovery, custody and next action

source.json records the exact source commit, relevant tracked/new binary patches, tested driver recovery and existing media references. The retained executing page/guest module is e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467; the locked rebuild has the same hash. All 647 runtime pins matched after the run. Git and the recorded patches supply source recovery; no source tree or media archive was made.

Both fixture browsers closed. The production browser exited zero; its driver and Chrome PIDs 2311285/2311292 are absent, ports 37933/41189 are closed, profile/writer lock are absent, final buttons are zero and input queue is empty. Closing the page destroyed its observers. The conversion browser exited zero and its profile was removed. The owned messageboard watcher was closed. Sampled free disk remained above 2 GB.

Next, use the corrected retained driver for a separately authorized fresh Dungeon Keeper original. Include the new pointer module in staging pins, and recompute pin lengths/hashes for the changed browser and dependency. Preserve the predecessor's ordinary capture gesture and reviewed-capture admission; observe the live automatic capture decision. Require actual lock whenever intent is true. Qualify useful consumed motion before selecting Start New Game or claiming Eversmile, dig or build. Never resume or replay the predecessor's closed UNKNOWN lane. This fixture proves none of those game checkpoints.