Recoil orphan-safe native attempt, 2026-11-08

Run: scratch/runs/20261108-recoil-orphan-safe-native/.

Concrete blocker; gameplay goal remains incomplete. The sole fresh native original reached Software / name user / Medium, completed the natural opening, and visibly resumed from a personally reviewed current menu. The unchanged 4500 ms gate refused ordinary M because both actual HUD candidates exceeded 250 ms freshness. Zero M/C/arrows/attack DOWNs were dispatched. There was no replay, second original, invented repair, or fixture substitute for gameplay. No new substantive generic defect was established; this is not a completed M-consequence or power-control movement result.

Actual original evidence

closure.json verifies all 838 source/media/helper pins, the actual hermetic preloader and compound HUD dependencies, all 295 inherited artifact references, and the prior accepted controls by exact hash. The tested subreaper, lock, WNOHANG reaper, double-fork and failclosed controls were reused without reruns. The exact locked-launch-fixed.py, owned-custody.py, and reap-owned.py were uploaded and invoked; the only wrapper substitution is its owned lock name. custody-dependencies.json records these selected bytes. An initial preparation failure omitted a generated phase helper from the copy list; it was corrected before allocation and retained in initial-preparation-failure.json.

One UUID-idempotent allocation produced bx_rukvjz6t. Observable independent normal/watchdog processes were armed before creation; the immutable watchdog deadline was 12:11:31.018Z, browser deadline 12:08:31.018Z, overall cleanup 12:12:31.018Z. The remote custodian enabled/verified subreaper adoption and held /tmp/recoil-orphan-safe-native-runtime.lock before browser startup through complete descendant drainage. No local runtime lock was acquired. Free disk exceeded 5 GB before allocation. No old box was resumed/deleted, no deployment, agent or worker was launched.

The personally inspected captures separate phases:

The current reviewed PNG is 109dc596c99e258507b04fdbf87fe06463df533c8506dd0e6b6ede05be5938b4, CRC checked, 1024×768. Its visible guest cursor hotspot 459,249 lies inside Resume 380,247–504,279; Resume is highlighted. resume-current-png-proof.json binds these bytes, and request-ordinary-map.json records the personal review before the click. The historical 459,249 witness was not reused as authority.

map-receipt.json retains baseline plus five post-click probes. Early probes still show MENU. The two later 640×480 native images were personally reviewed: health 100, timer 00:00:02 then 00:00:03, changed tank/camera view, MENU absent. Their SHA-256 values are 98f054ddecfa2794a7de8e16a009efaf8dca81b116b737c41088934614feb03c and 3ca9737bb69ba674f800d50952e5fb0c792e427c7cff9ef3113948fae3c09f66. These establish Resume's visible effect, not an M consequence or qualified movement. The decoder's numeric timer/health fields count mask pixels; they are not the displayed timer/health values.

Candidate Complete probe Final co-located owner/pixels Decode Pixel age at check
timer 00:00:02 1530.39 ms 301.16 ms 13.16 ms 315 ms
timer 00:00:03 1977.74 ms 318.33 ms 8.62 ms 327 ms

The retained per-read trace also shows guest identity evaluations taking 718/909 ms and renderer identity evaluations taking 390/509 ms during these candidate reads. These measured spans locate the delay in target evaluation / final pixel response, rather than attributing it to decoding or serialized module reads. They do not distinguish target scheduling from pixel encoding inside the final response. Earlier MENU pixel responses took 15–18 ms.

Parallel guest/render Promise.allSettled, four complete uncached module responses per HUD probe, exact targets/instances/epochs, final co-located owner/pixels, 250 ms freshness, timer temporal pair, gate4500/M500 and independent release remain unchanged. Gate Escape DOWN occurred 1 ms after gate expiry; late probe completion was retained and could not send a gameplay DOWN. The transport acknowledged all 21 commands; that is not proof of native gameplay consumption. The after-map capture does not establish native MENU reopening. No input was repeated to compensate. The original bounded observer expired normally; one fresh bounded invocation of the exact unchanged passive helper preserved its earlier owner/parent records and observed the gate window.

Custody, retrieval, and shutdown

The repaired custodian worked on the actual native Chrome lifecycle. Its census retained adopted crashpad and other descendants by PID/start identity before reaping. Native waitpid records include six adopted Chrome descendants and the browser-driver child. verify-drain.py independently joined the pre-close full ownership census with the final custodian census: all 26 PID/start identities absent or reused, wrapper absent, remote lock independently acquirable. terminal-inventory.json also verifies both ports/endpoints closed, profile and writer absent, observed keyboard/mouse released and normal browser cleanup. No undocumented descendant exception or wrapper kill was needed.

Before normal stop, the immutable collector retrieved and independently reread 407 artifact references / 59,595,955 referenced bytes / 347 unique hashes. verification.json independently rechecks every reference again and validates all PNG chunks/CRCs/dimensions: 136 unique PNGs. Unique logs, captured modules, phase images, observer images, original failures and custody records are retained. Unknown/pending semantics remain explicit; nothing authorizes replay.

Final box bx_rukvjz6t: archived, expiry null, stopped 11:54:54.646Z. Normal stop stop_e20e9a26b23210d8 completed 11:54:54.681Z, forced=false, error=null. The independent watchdog and normal lifecycle both exited 0; supervisor reports no remaining children. The prior scheduled expiry was 12:12:31.205Z. Stop occurred promptly after verified retrieval, well before that expiry.

source-head.txt records checkout SHA; selected original source remains 49874a81c5ae180c20a15c6bb169d9041f028cae through recovery.json and the accepted handoff's relevant binary diff, small untracked patches and existing media path/SHA recovery. owned-inputs.patch records small run helpers only. No source tree/media archive or peer WIP is included. No production code changed; there is no production build or generic-repair claim.

Next: diagnose the retained active-target/final-pixel latency before any new original authorization. A future pass still needs qualified ordinary M and its visible consequence before reviewed ramp/control movement; attack requires a reviewed living target. Full27, parity, audio and FPS remain open.