AVP2 asynchronous key recovery

Evidence: scratch/runs/20261107-avp2-async-recovery/ (ignored).

Recovered the interrupted avp2-async-world-2154 edits. Its retained event stream ends after launching the local Worker regression and focused tests; there is no allocation or original boot command, allocation artifact, result, final message or release. The before/after local Chrome and driver PIDs and ports are absent/closed, and the predecessor board watcher is absent. predecessor-lifecycle.json and predecessor-cleanup-audit.json record this audit. The reported launcher/session termination was 143. No retained signal sender audit identifies its cause, or proves a cross-group SIGTERM. The launcher has no kill instruction. The inspected Puppeteer cleanup targets its owned browser PID group with SIGKILL, which does not establish the SIGTERM sender. Historical process-group membership is unknown.

The production guest Worker regression executes ordinary x86 imports of GetAsyncKeyState through the shipped host/Worker path. Before repair, trusted Enter DOWN/UP with an undrained queue produced 0x8001, 0x8000, despite physical release. After repair, that case and the drained case produce 1, 0; a held key produces 0x8001, 0x8000. The low bit remains a pressed-since-query latch, consumed once. GetAsyncKeyState now reads the existing physical key state for its high bit, independently of the queued message snapshot. Message ordering, modifier ownership, physical reset and per-query consumption remain covered by the focused 55 tests. The reviewed host-file census seal was refreshed; callback ownership remains pending.

The permanent browser regression records PID, parent, process group, session and start ticks and checks that Chrome owns a distinct cleanup group. Runs are launched in detached, precisely owned sessions; no substring/group kill of shared processes is used. A new locked build attempt stalled after the region allocator printed success. Its exact Node PID 2738355 was waiting in shutdown futexes; retained strace records this without attributing the earlier worker interruptions. Only that verified owned PID received SIGKILL and the build was rejected. The allocator exits normally with --v8-pool-size=1; the complete locked build passed with that host-tool option (build3-exit.json, exit 0). The focused 55 tests and two production browser Worker runs passed. Code commit 74d173af4 was immediately pulled and pushed on main.

Run controller release checks retain trusted DOWN/UP, independent UP, owner identity and sticky UNKNOWN. They retain _asyncKeys as diagnostic message state and use the non-consuming physical high-bit accessor for release. Neither an unconsumed low bit nor an older queued message means physical hold. The qualified Gate, independent-UP and entry-identity helpers remain unchanged.

One original Marine followthrough

The predecessor allocation/boot audit permitted one new UUID, bx_pcy2gh6p, created 2026-11-07 22:03:55 UTC with initial expiry 22:38:55 UTC. The detached observable watchdog's immutable 25-minute stop deadline was 22:28:55 UTC. Local and remote launch floors exceeded 5 GB. Runtime closure, 619 pins and 18 referenced media bodies were checked; the existing collector 8453f67b4 and qualified controls were reused. The actual compiled module remained f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121. No second original was allocated or booted.

Reviewed captures progress from Single Player to Marine / Unwelcome Guests / Normal, through the introductory conversation into a recognizable first-person corridor with a Marine teammate, knife, health 100 and armor 050. The loading screen was not evidence of a hung world. intro-progress, after-strafe and return-corridor show bounded D/A strafes changing static corridor geometry and returning to the corridor. All controller commands received SUCCESS; independent UP left no held input and no sticky UNKNOWN occurred.

Attack remains unproved. Both bounded mouse pulses produced trusted DOWN and UP and physical mask/high bit zero afterward. Retained event coordinates, however, are (0,0) with an empty target for both pulses. The separate native pointer move reached (350,450) on screen, but the long-lived Puppeteer controller's distinct Mouse object retained its own (0,0) coordinates. The post-pulse canvas-target assertion correctly failed, with a known ordinary assertion exit (not UNKNOWN). No attack, hit or damage claim is supported. No input was replayed after this discovery. The controller and original were closed ordinarily. A future authorized original must aim using the same controller Mouse object and verify its trusted target before claiming attack.

Passive rendering boundary

producer-original-draw-assets.json retains natural legacy producer evidence: 565 accepted draw API calls, 62 GPU submissions, four covering GPU draws, and zero texture references on sampled triangles. Sampled fixed-function stages select diffuse color, with empty texture arrays. Native GL fences passed with no fallbacks. The 91-entry VFS inventory was read through descriptors without invoking lazy data accessors; pinned HTTP bodies completed. These observations place the sampled plain-color output before texture sampling but do not identify a faulty texture creation/binding operation. No WAT/JS rendering fix is demonstrated, so none was made. The existing stage-zero SetTexture handler was inspected read-only.

An earlier passive inventory accidentally accessed a lazy .data accessor and raised a known VfsPendingError; its logs remain preserved. A subsequent read-only descriptor inventory recovered evidence without replaying input. The public legacy draw trace produced no usable supported call records and is not evidence that assets or draws were absent. The personal review text for loading-current originally mislabeled it as loading; the actual retained image shows the interior conversation, as subsequent reviewed captures confirm.

Evidence includes exact source SHA, relevant binary diffs/new-file patches, module hashes and existing media path/SHA references. No source or media archives were created. The terminal SIGTERM sender remains unknown; the MW3 worker's exit 143 does not establish a common cause.

Retrieval and delegated shutdown

All 142 inventoried unique artifacts (16,024,882 bytes) were retrieved and verified by hash and size before the normal-stop request. Exact owned driver, Chrome and late-controller processes were absent; ports 37637/35589 were closed, profile absent, and held input empty. The separately owned local board watcher was stopped only after its exact cmdline was verified. Both lifecycle delegates exited 0 with terminal records; supervisor reports no remaining child. The normal, non-forced stop stop_a87926f01499d3cb completed at 2026-11-07 22:18:14 UTC. Box bx_pcy2gh6p is ARCHIVED; its former expiry was 22:38:55 UTC and the current API expiry is null. Custody returns to the coordinator; do not resume or create another original under this task.

Next concrete work is a locally tested controller-coordinate fix before any separately authorized attack attempt, plus an upstream texture creation/binding trace that identifies an actual fault before changing rendering behavior.