Nox interrupted worker recovery: pointer boundary remains unobserved

Recovered the terminal nox-window-observed-staff-target-0140 run after its worker exited 143. The retained original remained on the main menu after its first Multiplay click. No further live game protocol, input, capture, retry, quit, extension, allocation, or resume occurred during recovery. No pointer implementation change is justified by the retained evidence; the concrete blocker is the missing DirectInput return and Nox cursor-consumer boundary.

Evidence: scratch/runs/20261110-nox-interrupted-pointer-recovery-0150/; recovered artifacts remain in scratch/runs/20261110-nox-window-observed-staff-target-0140/. Native source base: 938d8b6932ead8b338b8bbda22c6761a69ae941c, with that run's owned.patch, generated source patches, and phase seals. Recovery base: f3d9d78741bc8e39f381e15c1fe0a658f168c187. Git and small patches are the source archive. Other workers' changes were preserved; Q3 startup, shared providers and LBreakout were read only. The accepted fixture repair 682f77655 and readiness repair bce2b316f were reused unchanged.

Artifact recovery and lifecycle

The predecessor's event tail ends at its cleanup-verification/retrieval command, without a result or final. Recovery found its game processes already terminal. The retained cleanup record establishes separate stdout/stderr EOF, actual parent waits, complete kernel fork lineages and all descendants reaped: fixture 150, original 206. The original supervisor returned 1 following the owning SIGTERM cancellation. This does not explain why the outer worker exited 143.

retrieve.log, independent-recovered-artifacts.json, recovery-summary.json, and the predecessor's remote-inventory.json / retrieval-verified.json preserve 167 remote artifact references, 150 unique SHA-256 contents totaling 5,556,660 bytes. Retrieval deduplicated against existing local content, including the exact tested module. Each reference was independently reread and checked for SHA-256 and size before shutdown; all eight gzip references decompressed successfully and all five PNG references passed chunk CRC checks. Unique logs, captures and module evidence were recovered without source archives or media copies.

The expressly inherited box was bx_jpmn9aqq, bound by the creation record and independent initial state to UUID 68932531-99ab-4ae4-b56e-23f95cc149f3. Initial API expiry was 2026-11-10T02:08:36.271Z; the immutable watchdog deadline was 2026-11-10T02:08:33.947Z. Its original supervisor/watch PIDs were 3241875/3241876.

Only after the independent artifact checks passed, recovery wrote the inherited watchdog's normal-stop request. The watchdog made the sole unforced stop and its supervisor actually waited for PID 3241876: return 0, stdout EOF and stderr EOF. Independent API lookup and a separate stop-status lookup confirmed the exact box's completed, unforced receipt stop_54db1492aca02f7c, completed 2026-11-10T01:49:27.810Z, before both deadlines. Final state is archived (stopped), expiry null. No box was created, resumed, deleted, or left running. Lifecycle assertions are also in offline-boundary.json; raw lifecycle records remain in the predecessor run.

What the recovered pointer trace proves

Both original menu PNGs were individually inspected during recovery. Multiplay remains visible after the click; the game-drawn pointer stays at the upper-left edge. The retained fixture review reports trusted DOWN/UP, production renderer DOWN/UP, DI button records 1/2, final buttons zero and the model-reviewed target/crosshair. Fixture success does not establish original-game cursor agreement.

The exact original command was normalized canvas point [0.46,0.427], 150 ms hold, 1600 ms bounded wait, bound to step-001-title. Its geometry was canvas 640×480, CSS rect (0,34,945,855), physical guest viewport (0,73,945,709) within the canvas. The generated native helper requests client (434.7,399.085); Chromium's retained trusted DOM record is integral (434,399). The production browser bridge maps this to logical (293,204), then the production viewport transform maps to native (293,196). That lies at the visible Multiplay destination. This is not evidence of a driver using the wrong coordinate space.

offline-boundary.cjs executes the production _mapExclusiveInputPoint against those exact retained values and checks the recorded result. Source hashes for the bridge, renderer input, observer, driver and DirectInput implementation match the native plan (source-identity.json). The preceding title move initialized the host reference to (319,238) with no DI motion; Multiplay then queued X=-26, Y=-42, left DOWN and left UP, in order. Four FIFO words drained, with head/tail ending 6/6 and host button mask zero. Pointer lock was false. The renderer reference and final position agree at (293,196).

The observer deliberately records drained payloads as null: a head advance cannot establish what records the guest caller received or how it used them. The retained snapshot still has GetDeviceState accumulators (-26,-42) after buffered FIFO consumption; these are independent channels in the implementation. Neither this nor the first-move baseline establishes an original-game defect. Changing the baseline, injecting compensation, changing waits, or assigning the guest cursor would be speculative.

Missing boundary and implementable next step

The run does not retain the mouse device's active format/axis mode, actual GetDeviceData/GetDeviceState arguments and returns, or the original executable's subsequent cursor update/menu hit test. Static source permits several distinct outcomes: a drained word may be omitted by di_format_offset, reported as relative or accumulated absolute motion, discarded by the caller, or used by a software cursor whose position differs from the host's. None is selected by the current evidence.

A separately authorized experiment should prearm a bounded, read-only guest-side observation around the original mouse polling calls, before the first input. Record device identity, acquired state, data format/object offsets, axis mode, request flags/capacity, HRESULT, returned count and returned record bytes at the real call return. Bind these to the existing command/image identity and FIFO sequence. Follow the actual caller's cursor-coordinate writes and menu hit-test entry/return to identify where (293,196) diverges. Use finite records and the existing immutable deadline; preserve first-failure cancellation. This is an observation plan, not authorization for another live attempt in this task.

Validation

The offline retained-trace/production-transform and exact lifecycle assertions pass. flock /tmp/wine-build.lock bash tools/build.sh passes; primary module remains 3e491c0840bb7771328738a0d160b4bbaa3d7ab72074cbd38a7b41e5d7e8ec18, compatibility module ecb45716b1ee7fc5218b1670dd270b19c7664ee30d000b3360a12786a1c21eaf. Fresh inspection and generation pin 781 inputs and preserve the accepted driver, media and readiness behavior. All four fresh actual-generated fixture controls pass (success, owner mismatch, changed PNG, missing UP); the generated observer integration passes all 12 image bindings with separate EOFs. They use the offline Node CDP peer and production bridge/renderer, never local Chrome or a native guest. The repository evidence checker reports no containment errors for this recovery or its predecessor; its 23 unrelated global errors remain untouched.

No original Multiplay transition, Jack creation, bunker visit, staff impact or gameplay acceptance is claimed. The released Nox code/test files remain unchanged. This task ships the recovery report and exact missing boundary, then releases ownership and stops.