Nox reviewed physical cursor anchoring

The single authorized original observation reached Multiplay, Network, Host Game and the Conjurer selection screen. Default Jack and the bunker staff were not reached. A Worker Runtime.evaluate timeout ended the run; no second native attempt was made.

Evidence is in scratch/runs/20261110-nox-reviewed-pointer-anchor-0300/. The original used published anchoring change b35771df0. Its recovery base is 703fe8732, with exact owned.patch, explicit plans, generated provenance and module pins retained. offline-base.txt and offline-repair.patch describe the subsequent offline repair. Git holds the source archive. No source archives or media copies were added to evidence. Shared renderer, startup, CRT, Q3 and LBreakout providers remained read-only.

Explicit anchoring and visible progress

The Nox-only observation driver preserves the befbafc5b button synchrony guard. An unprimed ordinary move now requires an explicit anchorOriginalCursor declaration. Admission checks the observed original cursor, the requested physical point through the observed canvas/viewport, and the absence of button effects. An existing unequal reference refuses. The command issues one ordinary native move, returns its own image, and checks that the original cursor stayed unchanged while the renderer reference became equal to it. The normal image review gate precedes dependent input. There is no inserted motion or guest write.

All nine images were individually model reviewed. Ordinary Escape advanced the splash to the main menu. Passive original calls showed cursor (0,0) and the pointer observer showed a null reference. The separately reviewed move at that original cursor established reference (0,0) without returning any axis or button data. Its returned image was reviewed before the Multiplay click. That click visibly changed the menu; original cursor and hit tests reached (300,195).

A capture-only observation showed the settled Westwood Online/Network menu. Network opened the map browser, and Host Game produced the native flag placement prompt. Host Game also repositioned the original cursor to (408,239) while the renderer reference remained (115,62). No button was sent while those positions differed. A separately reviewed ordinary move to the lower-right canvas edge returned original DI axis records (0,524) and (4,417). Original cursor, hit tests and renderer reference then agreed at the native bounded corner (639,479). Its image was reviewed before map placement.

Map placement reached a legible Conjurer selection image with Accept visible and the other two classes marked unavailable in the demo. This image was saved before the terminal failure and reviewed from retrieved evidence after shutdown. No input followed it.

native-consumption-analysis.json binds seven input commands to their prior image hashes and actual Worker identity, with 8,150 retained passive records. Every input interval includes original DI returns, cursor observations and natural hit-test calls. The capture-only settle-multiplay command had pointer and image binding but did not freshly arm an original passive interval. That is a concrete evidence gap, repaired offline below; this run does not claim complete per-command passive coverage.

Terminal failure and offline repair

After the final PNG, capture binding, state and pointer snapshot were saved, the cumulative passive finish evaluation did not return. The driver recorded CDP UNKNOWN Runtime.evaluate; no replay at 03:13:07.536Z. The independent owning watcher sent SIGTERM at 03:13:07.552Z. Custody reaped 356 of 356 owned descendants with complete fork lineage, actual parent return 1 and separate stdout/stderr EOF. Owned processes and ports were absent. No later game protocol, input, capture, retry, quit or extension occurred.

A file-only evidence fetch was started before the map-placement ACK returned; it found the missing final consumption file and failed. It did not issue game protocol or create a capture, and was not retried. Future operation should wait for a successful ACK before fetching a completed image bundle.

The retained passive records serialize to 19,675,640 compact JSON bytes for the session, versus 3,358,163 for the final input interval. The timeout's root cause is not proven. The same-worker offline repair makes finish return only its command interval, with global sequence/count metadata; full retained history and the original global record/event/time bounds remain unchanged. Capture-only commands now arm their own original passive interval before waiting. Neither repair has a second native validation in this task.

Validation

Before allocation, the generated original real Worker handler/installer and actual host initialization passed canonical-plus-observation installed-instance admission. Sixteen emitted move/click/anchor controls included firstMove, refMismatch, wrong anchor, anchor nonresponse and sticky refusal with no later effects. All eighteen identity mutations refused. Four production pointer controls passed, including the original first-move failure and an explicit anchor followed by actual compiled relative-axis drain. The locked build passed. Accepted fixture 682f77655 and readiness repair bce2b316f were reused unchanged, without a native fixture repeat.

The legacy generated driver sweep passed all 27 controls. The subsequent offline repair adds capture-only success/refusal and compiled command-interval isolation controls. Its focused controls and another locked build pass; final emitted controls cover the repaired source closure. Production primary module remains SHA-256 4f2c69a1528e01a9ab5127201d5e8a222da2a35236377a1351f444f55fa5a7d9, 2,186,099 bytes. The installed native observation module was 716d28c256b2d204e83a3e73a3bcd843e1afab9d62c957f868b762faacef19f5, 2,186,872 bytes.

Lifecycle

Only bx_8jfc36ud was created, UUID 893794fa-b8cd-4843-8ffe-8164fea88cd9. An observable independent watchdog was prearmed before allocation with immutable deadline 2026-11-10T03:32:07.743Z; API expiry was 03:32:08.248Z.

Before the sole delegated normal stop, independent local rereads verified 145 artifact references, 126 unique contents, 137,805,708 unique bytes, four finalized gzip streams and nine PNGs. PNG checks include chunk CRCs, complete zlib inflation, dimensions and scanline bounds. Actual owning parent wait and separate EOF were verified before stop authorization.

Unforced receipt stop_c1e6380d748451dc completed at 2026-11-10T03:14:08.819Z. Independent API and receipt lookups confirm archived, expiry null and forced false. The watchdog supervisor actually waited, return 0, with separate stdout/stderr EOF; both watchdog processes are absent. No box was resumed or deleted, and preexisting boxes were untouched. There was no local Chrome, nested worker or deadline change. Home free space stayed above 1.5 GB.

Next work requires separate authorization for an original observation with the smaller command-scoped passive return, then Conjurer Accept, default Jack and bunker staff. This task releases ownership and stops.