Recoil reviewed visible Resume, 2026-11-08

Run: scratch/runs/20261108-recoil-visible-resume-objective/.

New visible result: the personally reviewed, guest-drawn cursor was placed inside the actual highlighted Resume button before the sole Resume command. Native MENU visibly closed. Two independent passive captures then show the living HUD, health 100, and timer 00:00:04 → 00:00:05. Ordinary M and reviewed ramp/control movement remain unobserved: the unchanged 4,500 ms gate sent zero M/C/arrow DOWNs, because its next HUD probe finished after the deadline. No Resume retry, gameplay input, attack or second original followed.

The handoff at b5d322cf4 was read and hash-matched. Before allocation, all 806 providers in the composed manifest passed exact size/SHA-256 checks: the prior 790-file position closure plus sixteen owned run helpers. The actual hermetic preloader, explicit manifest, real HUD decoder, visible-pointer helper and accepted explicit 500 ms Escape helper loaded with an intentionally wrong ambient PIN_FILE overridden. All eight inherited control records and four position qualification/preflight records were hash-checked and reused, with no standalone control reruns. Remote launch checked all 806 pins again after 807 individual uploads, including the manifest.

Selected game source is 49874a81c5ae180c20a15c6bb169d9041f028cae; preparation checkout is 1e096bb120ad2cc0f3500cb77faf72657bef5b2a. Actual guest and owning renderer module SHA-256 are both 6cf120d0d3b3db1823a8105eb26ef52a9093a7f871c37a46117475e213fda1c0. The prior four position helpers were copied into this owned run; the new manifest explicitly pins the owned browser and dependencies. Peer WIP was excluded. source-identity.json, relevant-source.patch, new-lib_wininit-boot.js.patch, owned-inputs.patch and media-recovery.json record exact Git identity, relevant binary diff, small owned helper patches, and existing media paths/hashes. No source archive, repeated media copy or production edit was made.

Native Software/OK, Game > Start, NEW GAME and START were personally reviewed. The name remains user, difficulty Medium. Loading shows Island Storm and briefing 1.1 Operation Short Circuit; passive observation shows 1.4 Go to Landing Site without briefing input. Opening camera, tank transport, disembark view, living HUD and native MENU are separate observed phases. The first living HUD capture shows timer 00:00:02, health 100 and the tank facing the landing ramp. No interval is described as an exact simulation pause or exact active duration. The original read-only ready poll expired before startup finished; its failure and later ready/startup evidence are retained. No launcher or input was replayed.

The inherited Escape request explicitly holds 500 ms, independently releases, and produces visible native MENU. In that screenshot the actual cursor is below Resume. One ordinary relative pointer move [80,-160] while MENU is displayed produces review/resume-check-one.png. I personally reviewed the actual arrow tip [459,249] inside the actual highlighted Resume interior [380,247,504,279]. Host virtual coordinates were not placement evidence.

Before Resume effects, resume-witness.json records and validates PNG CRC, 1024×768, review reference resume-check-one, that hotspot/rectangle, and exact screenshot SHA-256: 9df9d8c3fcad922b98e5b993f7d2428c73483537e4a0434466f794c67da1b9ee. reviewed-resume-command.json carries that witness through the owned browser's pointer admission guard. No unreviewed cursor inference was substituted.

The sole ordinary-map command preserves m500 and gateMs4500. Compound start is 11:04:55.957Z, gate deadline 11:05:00.457Z. It performs one Resume mouse DOWN/UP. The five probes comprise the MENU baseline, three further MENU images and one fresh HUD candidate whose image starts 11:05:00.742Z and inspection finishes 11:05:00.932Z, after the immutable gate deadline. No temporal HUD pair qualifies inside the gate. The prearmed close releases input and sends Escape at the original deadline. No gate extension occurs. The exact source of the long final probe is not established; there is no claim that changing serialization would suffice, and no speculative timing repair.

The unchanged bound-input.cjs and compound-v4.cjs execute two complete uncached guest/render module byte checks per HUD probe, exact instance/epoch joins, final co-located page-owner references and pixels, 250 ms freshness and the timer-mask temporal pair. The late candidate remains inadmissible despite its fresh pixels. There are 21 commands, 21 exact ACKs, zero command UNKNOWNs, no M/C/arrow DOWNs and no attack. The generic cleanup template's literal pending UNKNOWN text is not a command outcome; numbered receipts establish the outcomes.

The passive observer independently captured native Resume's visible consequence before the closing Escape:

Both images were personally reviewed and CRC-decoded. Both show the living tank HUD facing the landing ramp; sampled M/C/arrows are released. This establishes visible native Resume and advancing displayed timer, not ordinary M consumption, player-controlled ramp movement, simulation FPS, or exact active intervals.

Two newly exposed generic helper defects were repaired and focused-tested in the owned run; neither repair was replayed against the original:

  1. Unreaped owned child retained the remote lock. The startup/runtime wrapper correctly held its own remote lock, but omitted waitpid on its exited Node child. After clean browser shutdown the wrapper remained with one zombie. That failure is preserved. With game cleanup complete and only the wrapper plus its zombie remaining, SIGINT made the wrapper execute its finally block, release the remote lock and exit. An independent check then verified all 30 recorded owned process identities absent before stop. reap-owned.py and locked-launch-fixed.py add nonblocking waitpid(WNOHANG) before census/drain. test-reap.py reproduces the real zombie, reaps it with exit status preserved, leaves a live child untouched and propagates unexpected errors. test-custody-integration.py exercises the actual fixed wrapper with a child outliving its launch function, verifies recorded reaping, drain and lock release. Initial test-fixture failures are retained. These tests launch no game or browser and do not need the local heavy-runtime lock.
  2. Mutable local log aliases invalidated retrieved artifact references. Collection initially hash-matched three zero-byte remote stdout files to collect.log. Its independent pre-stop reread passed, then that log grew to 71 bytes when collection printed its result. Final verification caught the mismatch. artifact-alias-failure-and-repair.json and the failed verifier output preserve it. The exact zero-length bytes, whose size and empty SHA-256 are unambiguous, are retained in owned content storage; no remote replay is needed. artifact-custody.py and the corrected collect.py retain known owned bytes by hash independently of mutable aliases. Focused tests reproduce source mutation, exact deduplication, wrong-byte rejection and corrupt-existing-object rejection. This is a tested collector candidate, not production integration.

Before quit, continuous descendant census plus the current parent closure, exact owned paths, explicit helper roots and Chrome's referenced crashpad PIDs recorded the complete observed owned process set. After browser cleanup and the wrapper correction, an independent per-PID/start-time check verified all 30 identities absent, including custody, launch, observers and browser descendants. HTTP/CDP ports were closed, profile and writer absent, numbered commands drained, and last observed inputs released. Earlier readiness, unreaped-child, test and artifact-alias failures remain in the run. This closes the prior task's missing startup-lock and individual-descendant-check limitations without relabeling the wrapper recovery as a clean spontaneous exit.

All 295 remote references / 31,487,238 bytes / 244 unique hashes were matched or retrieved and independently reread before delegated normal stop. Final verification additionally checks the corrected stable artifact references and all 49 remote PNG references with CRC decoding. The post-stop mutable-alias failure and exact empty-object repair are explicitly retained above.

The sole UUID-idempotent box is bx_afvq3a5t, created 10:56:31Z, initial expiry 11:22:31.314Z. Observable independent normal/watchdog processes were armed before allocation; immutable watchdog deadline is 11:21:31.128Z. Startup and runtime held the own remote lock. No local /tmp/wine-runtime.lock was held for remote provisioning, supervision, retrieval or thinking; no local heavy runtime ran. After verified retrieval and owned process drainage, delegated normal stop archives the box at 11:07:14.246Z, final expiry null. Receipt stop_e067c3dd2575e338 completes 11:07:14.264Z, forced false, error null. Both lifecycle children exit 0; the supervisor records no remaining children. The final API inspection independently confirms archival. No box resume/delete, public deployment or subworker was used. Launch disk exceeded 5 GB.

verification.json, the numbered receipts, exact Resume witness, passive images, focused test results and retrieval/custody/stop records contain the evidence. Only this note is published; original and corrected helper candidates remain in the owned ignored run. No production code/build claim is made. Next work needs bounded passive timing evidence for the native probe that crossed the gate, while preserving every exact check and the unchanged deadline, before another separately authorized original M attempt. Reviewed movement requires qualified ordinary M and its visible consequence first. Full27, parity, audio and FPS stay open.