Recoil manifest-bound native objective, 2026-11-08

Run: scratch/runs/20261108-recoil-manifest-bound-objective/.

This continuation authorizes one fresh original with the manifest-bound launch-v2.py from predecessor 0d8587810. The prior pre-Chrome failure is closed. The seven timing/native-image controls, six full-browser controls and focused actual-browser launcher fixture are reused by path, size and SHA-256; none of those accepted suites is repeated.

Before allocation, every one of the 757 pins was read from its exact provider and checked for size and SHA-256. Published source is 49874a81c5ae180c20a15c6bb169d9041f028cae; module SHA-256 is 6cf120d0d3b3db1823a8105eb26ef52a9093a7f871c37a46117475e213fda1c0. Peer WIP is excluded. The unchanged corrected launcher selects explicit PIN_FILE and preflights the real hermetic preloader plus transitive HUD decoder with the same environment construction used for execution. A deliberately wrong ambient manifest is overridden. closure-check.json retains these checks.

Exact Git identity, relevant binary diff, small owned-input patches, retained Wininit patch and existing media path/hash references provide recovery. No source or media archive is created. No production source change or build is claimed.

The one UUID-idempotent allocation is bx_ha493bhx, created at 09:44:28.175Z, initially expiring at 10:10:28.022Z. Independent normal-stop and 25-minute watchdog processes were observably armed before allocation. The immutable watchdog deadline is 10:09:27.827Z. All 757 remote pins pass again; the launcher starts its only original Node PID 8178 at 09:44:49.207Z after the real remote preflight passes. Local heavy execution holds wine-runtime.lock.

The browser reaches native Software selection, Start, NEW GAME, user and Medium, then ordinary START. Relative native pointer reads and reviewed corrections are retained. Native 1.1 Operation Short Circuit is visible during loading; the predecessor's 1.1 power-control/map versus 1.3 rocket distinction remains settled. No briefing cycling input is sent. Loading, briefing, map and inventory are not pause.

Native outcome and first demonstrated integration boundary

review/living-check.png shows health 100, timer 00:00:03, and the tank on the beach route at 09:52:40.266Z. Ordinary Escape is independently released; native-pause.png visibly shows native MENU at 09:52:57.791Z. Reviewed relative movement highlights Resume in resume-ready.png at 09:53:21.165Z.

Command 25, ordinary-hud-gated, is the only compound attempt. Resume DOWN is 09:53:36.301Z and UP 09:53:36.442Z. The unchanged compound-v3/browser-v4 retain separate 4,500-ms transition and bounded command windows, immutable overall deadline, <=250-ms freshness and the temporal pair. The first HUD candidate's image starts at 09:53:40.180Z; its post-image identity/byte validation completes at 09:53:42.019Z, 1,839 ms later. It is stale and cannot admit a key. The independent gate timer sends Escape at 09:53:40.802Z, UP 81 ms later. Receipt actionDue is null, gateQualified false, and attempted empty. The command has an exact ACK, which acknowledges completion of the bounded attempt, not native ordinary consumption. All 27 commands have exact ACKs; zero UNKNOWN. There is no repeated Resume, second original, gameplay C/M/arrow DOWN, attack, guest write, private handler or forced scene/configuration/clock/seed.

The independent observer captures returned living HUD health 100, timer 00:00:05 at 09:53:38.506Z, and actual native MENU at 09:53:42.464Z. review/ordinary-consequence.png independently captures that final native MENU. The native-menu interval from its first visible capture to Resume DOWN spans 38.510 wall seconds. The enclosing living readings span 58.240 seconds while the displayed timer advances two seconds; this supports excluding menu time, without measuring exact actor freeze or exact active duration. The observed post-Resume HUD interval through 09:53:41.183Z spans 2.677 seconds with the same displayed 00:00:05. These coarse timer readings are not simulation FPS. Briefing/loading/map/inventory time is never counted as paused time.

Tested generic snapshot correction

The newly reached boundary demonstrates that exact module/owner validation can consume the freshness lifetime of a returned HUD image. The original launcher, 757-pin manifest, compound and browser remain unchanged. New run-local bound-input.cjs / compound-v4.cjs move pixel acquisition into the final validated page snapshot. Each probe still performs two complete identity reads, including actual guest and renderer byte retrieval/hashing, exact Worker instance/renderer epoch checks and final page-reference comparisons. Nothing is cached between probes. The final page references and canvas are read within one synchronous page evaluation; its conservative host-start timestamp includes that evaluation and image encoding latency. Slow final encoding still fails the 250-ms limit. No input, freshness, temporal-pair, deadline, release or UNKNOWN rule is relaxed.

make-correction.py reconstructs these small run inputs. correction.patch records the focused changes against retained inputs; owned-inputs.patch recovers the small setup and focused check scripts. The new local fixture manifest is distinct from the unchanged original's 757 pins.

snapshot-browser-qualification.json tests the actual full browser entry with a new generated source-read-stall fixture and retained native image assets. Each actual Debugger.getScriptSource response is delayed 750 ms; bytes and identity responses remain real. The unchanged probe reproduces stale HUD ages of 779–808 ms and admits zero gameplay. The corrected probe qualifies a fresh native-image temporal pair at 3,200 ms, age 33 ms, within the unchanged 4,500-ms gate. Trusted generated C DOWN is at 3,234 ms, UP 81 ms later; Up DOWN at 3,317 ms, UP 2,001 ms later. Overall Escape is independently released. These are generated fixture events, never native Recoil consumption or route evidence.

snapshot-identity-qualification.json passes seven focused new snapshot cases: changed final page owner, guest instance, renderer epoch, actual module bytes and detached Worker all reject; the co-located temporal pair qualifies only after 12 full guest/renderer source reads across its three probes; 308-ms final image encoding remains stale. The accepted predecessor seven/six suites and focused launcher fixture are reused without repetition. Both new browser processes exit 0; Chrome PIDs, HTTP/CDP ports, profiles and writers are drained.

Retrieval, shutdown and closure

All 346 remote file references / 33,820,191 bytes / 281 unique SHA-256 values are matched to existing retained inputs or uniquely retrieved, then independently read and rehashed before normal shutdown. All 55 PNGs pass CRC decoding. The initial immediate post-quit inventory sees the launcher PID still draining and fails before collection. That diagnostic is preserved. A distinct later read-only custody check confirms all owned launcher/Chrome/observer PIDs absent, ports closed, profile/writer absent and released inputs before retrieval/stop. No input is replayed because of that diagnostic.

Normal delegated stop archives bx_ha493bhx at 09:55:39.913Z, expiry null. Receipt stop_f35c84a11147c5a6 completes at 09:55:39.932Z, forced false, error null. final-stop-receipt.json independently inspects that state. Both normal/watchdog children exit 0 and the supervisor ends with no children. Local fixture PIDs/ports/profiles/writers are drained and the runtime lock is released. Disk remains above 5 GB. No old box is resumed/deleted, no public deployment or subworker is used, and peer WIP is excluded from this documentation-only commit.

The task closes on the substantive tested generic snapshot integration repair. Native ordinary C/M consumption and reviewed ramp/control movement remain open; no actionable target was qualified and no attack occurred. Another original requires fresh authorization. Objective completion, full27, parity, audio and simulation/game FPS remain open.

Owned evidence containment, required fields and artifact existence pass. The repository-wide checker retains seven unrelated containment errors in older peer runs; they are preserved rather than edited here. verification.json checks the original receipts, hashes/PNG CRCs, cleanup and focused corrections.