Recoil first-observed EXIT repair, 2026-11-08
The sole fresh original reached Chrome and passed the repaired independent producer's first challenged census, then failed shared-kernel custody before original-game navigation. The first new substantive defect is reproduced and repaired; this task stops at that permitted condition. Current native Resume, Software before/after, M and bounded movement remain unperformed.
Evidence R: scratch/runs/20261108-recoil-software-ready-1650/.
Starting Git SHA: 50896dcd56958f3f5d971ba3448a542069613cec.
Recovery uses Git, R/owned-code.patch,
R/owned-operational-inputs.patch and R/owned-new-inputs.patch, plus existing
media paths/hashes in R/position-runtime-pins.json. No source/media archives or
peer WIP were collected.
Normal tools/native-runtime-generation.py selects the prior
20261108-recoil-original-ready-1621/repaired-ready-runtime-pins.json and adds
29 relocated physical helpers: 1,037 providers. The prior 1,008 rows, including
the original 851, remain unchanged. Actual kernel selection was published
275dd018aad23b8e2988a78033cddaa2e25d8df3; bounded endpoint remains
11683651…; independent producer remains 68ea2f56…. Decoder remains
204938 bytes / f468ffce84da0f517df009682c111bbcb955b67f8b5a122350abbe5abc49e1d8;
compositor remains 227919 bytes /
cb21daaf7c0ce2b2cf80436c192bda4ed78930b1b07eaf21a20eb9adb7ce7b51.
The shared supervisor and normal helper closure were used without bypass.
Before allocation, the cheap production provisioner resolved all providers, joined 1,038 callback transfers and checked nested helper closure. Hermetic preloader/decoder import passed without Chrome. Actual remote provisioning then joined 1,038 uploads, content-verified 1,037 providers and published the normal barrier. Queue recheck and actual parent admission passed before atomic consumption of the sole original request. Previously accepted 2,717 evidence references / 750,433,471 bytes were independently rehashed in place; unchanged Chrome, owner, HTTP, Software and watchdog fixtures were not rerun.
The prearmed persistent producer worked in the actual runtime. PID5943 / start8503 was created before Chrome PID5979 / start8965. Its first fresh random challenge response bound driver PID5930 / start8213, actual boot/namespaces, Chrome profile/endpoint and listener FDs for ports43985 and34163. Producer startup took about4.3s; the fresh response took about3.3ms. This is actual startup proof, not Software qualification or a claim about game health.
At journal sequence1031, 16:50:27.328Z, the kernel supervisor refused
unexpected unlinked task PID6763 with real WNOWAIT CLD_TRAPPED status1541
(PTRACE_EVENT_EXIT/SIGTRAP). No earlier row registers that PID. The task's
start identity and exit message were not recorded before this assertion;
the failure does not establish its parent, TGID or termination cause. The
existing terminal repair only admitted EXIT after initial autoattach
registration. An initial SIGSTOP can be superseded before first observation.
tools/native_kernel_custody.py now admits only initial autoattach SIGSTOP or
the exact kernel EXIT stop as an unregistered task's first observation. It
records /proc PID/start identity before waitpid. The existing retirement path
still requires the consumed EXIT stop and valid GETEVENTMSG terminal status,
fsyncs its unresolved-lineage receipt before CONT, and never admits unlinked
userspace. Other unexpected stops still refuse. Delayed fork lineage uses
the retained identity and still refuses observed PID reuse. No private
supervisor, deadline change or broader lineage guarantee was introduced.
R/first-exit-before reproduces the exact failure with real kernel statuses: hold a parent's actual fork event, observe the child's initial stop in the test scheduler, send SIGKILL before exposing it to the supervisor, then expose the actual EXIT stop first. No wait status or GETEVENTMSG response is invented. The isolated failure wrapper reaps its fixture descendants. R/first-exit-after passes with the repair: no child userspace marker, retained identity before consume, durable retirement before CONT, reap before delayed parent lineage, exit0 and empty drain. All nine focused kernel tests pass, including prior SIGTERM/clone retirement, late-lineage PID-reuse refusal and WNOHANG-zero controls. The locked full build passes in R/build-final.log.
R/repaired-ready-runtime-pins.json selects the repaired physical kernel through the same normal generation mechanism. All 1,037 providers and 29 helpers resolve and pass nested closure validation; original 851 rows are unchanged. The actual-run manifest remains immutable. No second runtime was attempted.
The one UUID-idempotent box was bx_y26z5y6n. Its ID and initial expiry 2026-11-08T17:14:52.708Z were durable before runtime. Independent observable 25-minute watchdog roles were prearmed before allocation. Immutable browser and outer stop deadlines were 17:12:22.542Z and 17:13:52.542Z. The remote runtime lock covered startup until supervisor failure; the local heavy runtime lock was never needed or held.
The failed kernel drain is retained honestly: exitNULL, 133 owned /22 reaped,
111 remaining at failure, and one unresolved lineage. EXITKILL retired runtime
processes after supervisor failure. R/failure-runtime-drain.json independently
checks all134 recorded PID/start identities absent, PID6763 absent without a
start-identity claim, both ports closed, no runtime cmdline/FD writers and
runtime lock free. Only after that audit was the owned profile removed.
No input was sent or held. This is a failure cleanup audit, not a successful
normal kernel drain. CLONE_UNTRACED and nonleader-exec limitations remain open;
actual kernelForkLineageComplete:false is preserved.
All32 unique outputs /10,130,744 bytes were retrieved and independently reopened/hash-verified before delegated normal API stop. Final retrieval references the already retained identical files without duplicate downloads. The box is ARCHIVED, current expiryNULL, stop receipt stop_d969cf48280154d5, completed 2026-11-08T16:51:25.557Z, forcedFALSE/errorNULL. Both local lifecycle roles exit0 and their supervisor records no remaining children. Local fixture and watcher identities are audited separately in R/local-identity-drain.json.
Runtime launch to kernel failure was about56.1s (driver execution about53.2s). Original guest loading, MENU and active gameplay were not entered. There are no current Resume pixels, Software before/after qualification, M, movement, observed health or game timer. The historical Software196457/health100/timer09 image is not current proof. Six complete uncached WASM/JS responses through allSettled, current contexts/pipeline owner, fresh challenged independent census, final PNG freshness through decode/save at250ms, changed timer/living pair inside the immutable five-second transition, argument-free post-Resume proof and sticky UNKNOWN retirement remain required and unchanged.
Next is a separately assigned fresh original with this repaired kernel selected normally. Personally review current native Resume before clicking; qualify actual Software before/after before M; review native M's consequence before bounded independently released movement. Full27, parity, audio and FPS remain open. No old-box resume/delete, public deployment, private guest writes, config/clock/seed changes, replay or subworkers were used. RELEASE and stop.