Guarded CLI consumption and NFS native admission, 2026-11-13

The missing opt-in CLI consumption guard is implemented and qualified against actual test/run.js, including its real renderer and D3DIM Node Worker. Native NFS III race, Pause/Continue and driving remain UNPROVED: the sole post-qualification native admission sample was 1,488,797,696 bytes available on /home, below the strict 1,500,000,000-byte requirement. No NFS original was launched and no renderer selection or driving input was submitted.

Evidence is scratch/runs/20261113-cli-consumption-guard-nfs-native-0910/. The distinct remaining blocker is capacity after the consumer defect was fixed. The designated resource0900 worker's actual outcome was NO_ADMITTED_MATERIAL_CLOSED_STDOUT_POOL, with zero recovery. That result, verification and publication are referenced by path/SHA in dependency-observation.json. No other cleanup, floor polling, RAM gate or protected Pawn/LMarbles/recovery-data modification occurred.

Protocol and consumption boundary

{action:"guarded-bind"} returns an opaque owner, epoch and identity. These identify actual in-process instance/exports, memory/shared buffer, renderer/input queue, native window/canvas/dimensions, input handlers and, when present, the production D3DIM device/encoder, adapter, module, control and command buffers. No absent guest-main Worker role is invented.

Submit one supported edge using:

{"action":"guarded-input","cmd":"keydown:38","guard":{"owner":"<bound>","epoch":1,"identity":"<bound>","nonce":"<unique reviewed intent>","deadline":1794536000000,"reviewed":true}}

The deadline must be a fresh absolute wall-clock millisecond value, within the timer range. The example timestamp is illustrative, not reusable. Supported edges are keydown, keyup, mousemove, mousedown, mouseup; other actions and malformed arguments are refused. Ordinary legacy input commands retain their existing interface and behavior.

The envelope is captured into an immutable event/deadline with a monotonic expiry bound. Identity, health, deadline and the intent's one-use claim are checked immediately before the real renderer handler. They are checked again before real takeInput, which can otherwise reassert async DOWN state. Expired/replaced-owner pending effects are removed from the originally bound queue; cleanup releases survive. Nonce tombstones remain for the CLI process lifetime. Rebinding cancels the preceding epoch's pending input.

A frozen reply explicitly says queued:true,consumed:false. Query {action:"guarded-status",nonce:"..."} for its receipt; a consumed receipt means the physical handler returned, not guest dispatch or a rendered frame. After later expiry/cancellation, consumed:true preserves the fact that the handler previously ran. Effect/inspection errors are terminal UNKNOWN with noReplay:true. HTTP pending-command timeout cancels the guarded intent and returns a 504 carrying that UNKNOWN receipt; socket disconnection also cancels pending guarded work. Ordinary noncancellable promises retain their old timeout behavior.

{action:"guarded-release",owner:"<bound>"} is the independent release path. It cancels pending DOWN before it can execute and releases held keys/buttons through the originally bound renderer, without requiring a valid deadline or current owner identity. Native fault retention instead abandons guarded work and uses the existing host-only fault cleanup, preserving its prohibition on guest callbacks. Normal CLI shutdown releases guarded input.

Qualification and limits

before.log records the actual CLI regression test failing before the change: the new guarded action is unavailable. Parent198090574's actual legacy CLI fixture separately supplies the concrete expired/replaced-queue DOWN and known duplicate-nonce double-effect failures; its 16-control result was read first. This run does not relabel the parent's 1516ms Solitaire HELD frame as native NFS Pause evidence.

after.log and cli-test-result.json record actual CLI producer/consumer checks for queued-versus-consumed ACK, one-effect duplicate refusal, expiry between enqueue and consumption, long synchronous producer turn, actual queue/window/device replacement, inspection error, epoch replacement, independent HTTP release before pending DOWN, effect-error UNKNOWN, expiry with DOWN still in the real queue, physical cleanup and preserved UP, and ownership change immediately before actual takeInput. The separate timeout case uses the production transport and guard manager with a controlled effect counter, not a native game. It verifies that a timed-out ticket cannot later consume and remains UNKNOWN; it is not described as an actual CLI timeout.

The fixture's independent absolute deadline watchdog was observably armed before runtime creation and bound to PID/start ticks. The fixture and watchdog both have actual exit0 waits and stdout/stderr closure. The independent HTTP release subprocess also completed normally with a known receipt. No uncertain input was replayed. This is consumer qualification, not native Pause, race rendering or driving qualification.

regressions.json records the related existing control, stdin, frozen-control, input-state/clock, DirectInput mouse and fault-cleanup tests. Two stale test fixtures needed maintenance: the source-region fixture lacked real require and the scheduler binding and used an outdated trace marker; the tick test needed explicit --clock=batch after the default service-clock migration. Their original assertions remain. The locked tools/build.sh log is build.log. No broad historical fixture matrix was run.

freeze.json preserves the 133 original media and the prepared b8ac865a5 native catalogue/setup/export route by existing path/SHA, without copying media. Parent module 04c53419…f0a17, app registry and media pins remain unchanged; the CLI source pin deliberately changes. Current source identity is source-base-commit.txt, relevant-source.patch, owned-new.patch and helper patches. Git is the source archive; no source-tree/media archives were made.

Next requires sustainable authorized disk headroom, then the unchanged native catalogue/setup/chooser export transfer into nfs3_chosen_renderer. Personally review ordinary renderer selection/race and two settled native Pause/Continue frames with stable world timer before one reviewed >=1500ms interval, with independent UP/Escape and a HELD frame before UP. No automatic route, extra Present, forced call, guest/config/clock/AI edit or old UNKNOWN replay was used.

Box ID NONE, state NOT_CREATED, expiry null, stop receipt N/A. No agents, other Codex workers, public deployment or new boxes were created.