Reviewed native pointer coordinates, 17 November 2026
Evidence: scratch/runs/20261117-disciples2-reviewed-coordinate-space/.
The published validator compared a logical canvas point directly with a native USER rectangle. The actual renderer first refreshes its presentation, maps through the physical output viewport (when present), floors in guest screen space, and applies ClipCursor. Its ordinary drag/release path clamps letterbox points to the nearest source edge. Reviewed containment must not use that clamping to admit a point outside the reviewed image content.
The repair records the renderer-owned logical canvas dimensions, source HWND, source/destination transform, physical viewport/native rectangle, output size and filter availability, and native cursor clip with each retained complete image. Capture checks the same public context before and after encoding. Each finite intent compares this geometry with the retained capture. A caller's review object cannot replace the mapping. Target containment uses the mapped native screen point, including each child ancestor's actual native client rectangle. Invalid geometry, letterbox bars, clipped-away child pixels, canvas edges and a cursor clip that would move the intended point are refused.
The renderer still receives the original logical canvas coordinates exactly once. No input scaling, native queue receipt, native BUTTON witness, DialogBox identity, independent physical UP, nonce or freshness rule was weakened.
Reviewed captures now encode the actual complete displayed output. When a separate physical presentation filter owns that output, its full image and bars are normalized to logical canvas input units before PNG encoding. The hidden unscaled staging canvas is not substituted for the displayed image. A viewport without its actual displayed output is refused. The exact encoder bytes and original capture times remain retained by the existing registry.
An independent oracle executes production renderer input methods over 460 accepted points; 740 other points are refused across identity, scale, offset, physical DPR, letterbox, cursor clip and boundary cases. Additional units cover changed viewport, source HWND, output, window, owner, generation, actor, focus, stale evidence, forged review transforms and nonce reuse.
Authored native PE tests use ordinary ChangeDisplaySettingsA(CDS_FULLSCREEN)
and actual public window geometry. Host sizes 800x600, 1024x712, 1400x900 and
400x300 prove exact native coordinates via the returned queued WM_LBUTTONDOWN
lParam, followed by independently released UP and the guest's UP after DOWN
caption. At 1024x712, canvas (868,357) reaches native (700,300). Replaying the
previous published validator over this exact capture refuses the canvas point;
the patched production transport delivers the exact native point. This oracle
does not use the new transform helper to predict delivered coordinates.
A real native DialogBox child BUTTON under the actual presentation filter at physical 2048x1424 / logical 1024x712 produces its exact typed dispatch-return receipt and real dialog transition. A separately rendered partially clipped BUTTON is refused before its handler runs, with false consumption, followed by independent release. These are authored synthetic protocol attestations, not claims of personal review of an original game's UI. Selected full PNGs were also personally inspected after these tests.
The existing fourteen native DialogBox/real BUTTON/fault modes pass, including generation, retirement, actor, focus, semantic and stale refusals. Guard, input-owner, current-quest media/helper, reviewed-context and actual transport tests pass. Normal and early-release native delivery regressions pass. An additional 250 ms per-pump stress case delivered DOWN and UP but exceeded the 500 ms confirmation bound (1206 ms), so it is retained as a failed timing experiment, not qualification. The previously qualified 65 ms pump shape is tested separately and passed at 308 ms in the sealed replay (normal 135 ms, early release CANCELLED without DOWN). No timeout, confirmation or release limit was expanded.
VB explicitly released the whole scene, native USER helper and runner files after publication; integration began only after that release. No function-level concurrent edits were used. The completed repair passed the canonical flock build with actual exit zero and unchanged start/end source hashes. Earlier attempts encountered concurrently owned Neverhood QT census WIP or changing inputs; they are retained separately and do not replace the stable build.
The sealed seven-case mapping replay and normal/slow/early delivery regressions
use Git aa89c364bf37d716e26d2251c75f49e22436b4dc plus the exact owned patch in
prepublication-holder2/runtime-source.patch. Foreign WIP is excluded. Its
module SHA-256 is bc9ce60ee3c95a396f3d60f849b267d6fde84a890454afc6d4956cc5338f33e2.
All native/watchdog waits completed zero with both EOFs. A first sealed attempt
truthfully refused a missing test/helpers dependency before the child BUTTON
case; the helper source closure was corrected and the complete replay passed.
The holders were hash-verified and normally closed with actual successful waits.
No fresh original has started at this checkpoint. The authorized fresh original route is conditional on material publication and immutable native qualification.
The material repair is local main commit 78d99e63df492695c0355dd4ee1b318052468cf3.
A further Git-only sealed replay of that exact commit, with no working-tree
patches, passed all seven native cases (including the complete displayed-output
pixel assertion); actual native and watchdog waits were zero with both EOFs.
Its source SHA-256 is 78339c95363bea2dc8c714efe96dfd80cce7d01fd71c070c8f67804d50745f5e
and its module matches the sealed qualification above. One recipe initially
used the pretty-printed manifest file hash instead of the holder's canonical
manifest hash and was refused before any native/watchdog child; the corrected
recipe passed without a code change.
Publication is blocked: the required strict pull exits 2 because incoming origin/main changes would overwrite foreign WIP in the D3D worker, guest thread host, two GL tests, and callback ownership census/revisions. No foreign changes were stashed, reset, adopted or resolved, and no push was attempted. Owners and the coordinator were notified on the messageboard. The fresh original was not started: publication is an explicit prerequisite. There is no map action or native Quit claim. Next: reconcile those foreign files, publish the material commit, then qualify the exact published runtime before the one authorized original route. All task-created native processes and holders are closed; no box was created.