MDK2 ordinary post-Medium timer and deferred-load continuation

This assignment follows the ordinary Medium press, balanced release and active0 full rest established by 5e2542a569861d7458f336784b41012538d72926 and 5b72401a77b3afb5c57d93a77b305c1666d9a20e. That original stopped at3968, with command62 phase/raw/previous/active0, empty queue,384/512 aggregate menu steps and53/113commands. Its captured timer countdown was1.2000000476837158; no timer handler, demand load, new level or gameplay was witnessed.

Evidence directory: scratch/runs/20261105-mdk2-post-medium-timer-onward/. Git is the source archive: record exact Git SHA, relevant binary diff and new file patches. Existing original inputs, media and modules are referenced by path and SHA, never copied. No source snapshots or archives are created.

Preparation starts offline. Shared helper edits/full build wait for Return Fire0050 final runtime AND file release and independently checked exact normal cleanup. The actual original additionally waits for Stronghold0104 final remedy publication/file release and all builds to finish. No Stronghold or Return Fire files are edited. The runtime closure stays stable while Chrome is live.

One original local attempt is authorized, under exclusive /tmp/wine-local-runtime.lock, with no other Chrome or heavy local runtime. Before Chrome, the actual composed filesystem and HTTP route must pass plain Puppeteer and public/native SPAWN_FORBIDDEN preflight on the published binding. Historical manifests stay unchanged; necessary named amendments retain old/new SHA provenance. Existing35pins/502bytes/534translations per boundary, 50M/360s/512record observer,16ms frozen review, owner/byte/record/input/ACK and active0 full-rest guards,512/544menu steps and existing controller caps stay.

The existing ordinary author/operator/next/control path reproduces New Game, Medium and full rest, then advances only through freshly reviewed, natively qualified bounded continuation controls. No timer/render/load is forced; absent memory is unknown and queued input is not consumption. Every relied-upon fresh checkpoint is personally viewed. First post-Chrome fault stops the attempt; normal Browser.close and exact PID/startticks, ports, profile and flock cleanup precede offline diagnosis. There is no second original attempt.

Absolute Chrome cutoff: 2026-11-05T02:20:00Z; cleanup cutoff02:23Z. Existing720s browser,60s bootstrap,40s command and180s cleanup caps remain. Box ID/state/expiry: NA. Coordinator owns shutdown.

Return Fire final release 533e8faad7bee6f29e123acf3e9febe5b1b40b84 preceded shared edits. The independent check verifies Chrome1607987/start 188872167, Node1607950/start188872141 and flock1607949/start188872141 absent, HTTP40217 and DevTools36595 refused, exact profile absent and flock free. Its normal Browser.close protocol request and exit records bind code0/signalnull. An initial offline verifier incorrectly expected MDK lifecycle event names; that fixture failure is recorded and corrected before shared edits.

The only necessary shared amendment renews the local assignment,02:20/02:23 cutoffs and explicit ReturnFire0050 plus published Stronghold0104 release gates. It leaves the integration/resolver/observer byte-identical and does not amend any historical manifest or selected runtime row. Forty-five local tests and25 focused integration tests pass, including dual-release/build-order negatives and ordinary producer/review/input/trace guards. Stronghold signaled ALLBUILDSFINISHED at01:26:01Z and explicitly released build priority before the MDK full flock build. Actual Chrome still waits final file publication.

The full flock build passes and reproduces current e620f3ad/66b37bd7 modules. Stronghold final 7c73cca3087d661e6aba1f57df082970705fe129 and its01:27:13Z file release satisfy the explicit final publication/build gate.

Implementation d1c6c9796 was published by immediate origin and direct strict pinned SSH pull/push. Actual run Git was b00672c90e56a3c538e867064a1976f1412612d4; all relevant source differences were recorded before launch. amendment-ledger.json binds old/new local helper and test hashes, and unchanged integration/resolver/observer and module hashes. The actual composed preflight passed439FS/439proofHTTP/439actualHTTP and2253 installed-caller rows with plain Puppeteer mocked, public/native spawn forbidden, zero Chrome launches and both servers closed. Historical manifests and current runtime source/module rows were unchanged.

ONE original launched Chrome at01:29:40.526Z and reached New Game, then the difficulty screen with Medium highlighted at3648. Ten fresh PNG checkpoints were personally viewed, from startup0 through native-selector67; each review binds its PNG SHA. New Game needed48 steps, versus64 in the prior original. The ordinary Medium keydown was acknowledged at command76; its readback still had queued KEYDOWN13 and CHAR13, asyncPressPublished true, and native command62 raw0. This is queued input, not native press or consumption. No Medium full-rest, timer, loading or gameplay result is established by this attempt.

The first attempted Medium16-step command77 failed at01:32:54.463Z. The surfaced CDP error was TargetCloseError on Runtime.evaluate. The retained browser console records the underlying owning-Worker trap: gc return history missing at EIP40218c. Initial observer state at3648 had zero boundaries and only its initial sample. Its all-Worker frozen snapshot captured main PC449096, ESP074ffd68, saved44308f at stack index4, and saved40218c at index10 (ESP+40). Original static instructions show44308a calls449050 and44308f is its actual return. Together with the later missing-entry return, this supports an already in-flight GC when tracing attached. The earlier GC entry is not observed or reconstructed. The concrete prerequisite is native GC entry coverage before requiring that return. The task cannot progress by fabricating an entry or relaxing that guard.

The actual driver immediately took its first-command-failure cleanup. Normal Browser.close called01:32:54.494Z, Chrome1625277/start188938976 exited0/null at 01:32:54.606Z and closure resolved.607Z. No kill fallback ran. Independent verification at01:33:12.162Z confirms that Chrome, Node1624693/start188938366 and flock1624692/start188938366 are absent, proofHTTP40585/actualHTTP44711/ DevTools42721 refused, exact profile absent and external flock inode141907 free. A subsequent ordinary quit request found its owner process already absent; it submitted no additional command. No second original was attempted.

The necessary offline remedy preserves the first observer failure through the existing private run-unwind sentinel. Previously the observer's strict rejection threw through the ordinary guest Worker, closing its CDP session and losing the terminal report. The adapter now catches its own observer rejection at the same native run boundary already used for observation exhaustion. The observer still rejects missing GC history, stops tracing and becomes terminal. A later queued slice executes no native instructions. The report stays queryable, so the driver can persist the actual cause before its normal first-fault closure. Original native/import traps keep their ordinary handling. No observer byte, pin work, record/boundary/deadline guard, source/module manifest, guest state, input or controller cap is changed. This improves fault evidence; it does not establish GC entry coverage or authorize any further original.

The new actual shipped-Worker regression executes the original GC return boundary with an explicitly unobserved fixture prefix, independently for both current module variants. It fails against the prior adapter and passes with the remedy: the same missing-history cause remains terminal/queryable, no GC entry or return event is invented, and a second queued slice preserves all registers and the complete memory hash. Genuine original-import failures still trap for both modules. Twenty-nine focused integration tests and45 local tests pass. The initial new fixture used a duplicate local variable, then an incorrect expected status/error string; these offline fixture failures are retained, corrected and distinguished from the actual red regression. No fixture is an original gameplay attempt.

The prior two queued first-fault native regressions also pass: the adapter still stops before the rejected POP ESI for both historical module fixtures. The full flock build passes after the remedy and again reproduces the unchanged current e620f3ad/66b37bd7 pair. All builds are finished.

The diagnostic remedy is published as 3dafc52d07f404c42ede112e0e6c62d8725f9a20, with immediate origin and direct strict pinned SSH pull/push. Its exact published actual composed preflight again passes439FS/439proofHTTP/439actualHTTP/2253installed-caller checks, independent public/native spawn prohibitions, zero Chrome and both servers closed. remedy-preflight-result.json, remedy-amendment-ledger.json and remedy-relevant-diff.patch bind that change without source/module copies.

Final result: concrete GC-entry coverage prerequisite, with tested and published first-fault evidence preservation. The timer/deferred-load/gameplay objective is not established. Next work must prepare earlier ordinary trace attachment with complete input/owner/review/GC history, or independently establish a safe ordinary attachment boundary; retain the missing-entry rejection and all original caps. Any new original attempt needs separate authorization. This assignment stops here. Box ID/state/expiry remain NA; coordinator retains shutdown ownership.