MDK2 repaired launch: normal diagnostic and original construction progress

The authorized Chrome-only diagnostic passed with ordinary page work and normal closure. The single original then reached personally reviewed title, New Game, Medium and level1 construction. Its first input-free 16-slice construction chunk recorded 176 native boundaries without the earlier mid-poll assertion. It did not reach a completed resource, Lua execution, a constructed Level, or personally reviewed gameplay.

Evidence is scratch/runs/20261104-mdk2-repaired-launch-original/ (RUN). Accepted input ae7f8ed190c364bf60e3f650f89846d33dd355d4's exact 259-file, 692062-byte seal was verified unchanged. The accepted observer from 6f1b503a0 remains byte-identical; historical helpers, seals, modules and media were not edited.

Actual caller and diagnostic

launchOwned now connects the actual diagnostic and composed original callers to closeUnreturnedLaunch. A rejected launch immediately attempts normal closure of the single recorded child through its own late endpoint, retaining the original launch error and allowing no retry. The installed launcher still retains its own fallback behavior; no process signal is suppressed or reclassified.

The offline regression uses the installed BrowserLauncher.launch timeout producer and its actual child process implementation. An owned Node child emits a delayed endpoint. The old absent-Browser caller leaves it live; the repaired caller closes that same child with exit0/signal-null through a fixture connection. Identity, endpoint ambiguity, foreign endpoint, expiry, abnormal exit and duplicate cleanup negatives remain. This fixture is not a successful Chrome recovery claim.

Stronghold's explicit 19:15:07 local release was observed before the sole runtime flock was acquired. Diagnostic Node984038 spawned Chrome984080 at19:23:26.008Z. The caller allowed the realistic remaining 45-second launch window inside its 60-second browser cap, with a separate 60-second cleanup reserve. Launch returned, ordinary document/iframe work ran, the iframe was removed and Browser.close completed normally. Chrome exited0/signal-null at19:23:41.653Z; Node exited0. The diagnostic measured16403ms and recorded no local signal callers. Actual page/frame teardown, root disconnect and Browser.close are all retained. diagnostic-closure.json verifies both PIDs/process group absent, DevTools refused and the runtime flock available. There was no diagnostic HTTP server or guest.

Historical Chrome894623's termination attribution remains UNPROVED. The proved Puppeteer SIGKILL caller for the separate Chrome957916 failure remains unchanged. Neither is relabeled by this new normal diagnostic.

The single original

The dedicated adapter verified the current111-helper closure, four Git-pinned dependencies, both retained modules, all432 filesystem pins and all432 HTTP pins before the original release. RUN records the actual Git SHA, relevant dirty patch, candidate hashes, nonce, installed package hashes and existing module/media paths. The original release retained720s browser/60s bootstrap/40s command/180s cleanup bounds, with hard22:00Z Chrome and22:03Z recovery cutoffs. Its actual issue time was19:24:36.371Z, closing no later than19:36:36.371Z with recovery to19:39:36.371Z.

Every startup capture was personally viewed. The title/legal overlay was viewed before measured Escape. The readable highlighted New Game was viewed at3184 steps before its ordinary selection; the orange Medium in the difficulty dialog was viewed at3232 before its ordinary selection. Subsequent fade, black pending and construction captures were separately viewed before dependent chunks. No review was inferred from a prior run: this run needed one more opening observation than the earlier run before New Game was visibly ready.

At3328 steps the main readback had an absent Level and active type10 level1 request 6174784, buffer14713028, size192320. The before-first-script binding identifies the actual transferred/instantiated canonical module, main slot0/endpoint1/thread1; file Worker slot1/endpoint2/thread2 and unchanged frozen ownership also correlate. The3328→3344 quantum advanced file output from16384 to65536 bytes (+49152), with compressed input2048→10240. Main completion remained0 and completedBytes stayed 982692. Level remained tag-6/absent at both samples.

The176 delivered boundaries include the accepted UNKNOWN poll prefix, its exact Sleep-return suffix and subsequent fully observed poll returns. No earlier poll entry or demand call is invented. No completion-consume, finalizer, demand-return, buffer-call or Level-creation event was observed.

First continuation dependency and offline repair

The retained driver finishes the observer at the end of its sole16-slice quantum. The live report was terminal with status original-quantum-ended, while the resource was still incomplete. Its old consumed branch silently skipped tracing on later construction steps. Continuing would lose the history this task requires. The worker stopped here without submitting a second construction chunk.

The repaired driver rejects every further step after consuming its sole trace, before guest execution, including when a later review has a different scene. It does not restart the observer, extend its10000ms/100000-boundary/512-record caps, invent missing history or turn cursor progress into completion. The actual composed-command regression has old-fail/new-pass evidence: old starts1/finishes1/ steps2/no error; repaired starts1/finishes1/steps1 with native construction history ended; no unobserved continuation. ACK, ownership, review, deadline and producer negatives remain intact. A continuity-capable observer and earlier demand-call coverage are still required before another original can prove completion through Lua and Level. This change contains the gap; it does not provide that capability.

A second observation gap was also found: the original is compiled with its historical helper filename, which bypasses the passive preload's run-directory require filter. Spawn/exit/close/driver events were recorded, but original Browser/page method wrappers were absent. The caller now directly installs those owner-bound wrappers on successful launch. Browser/page attachment is idempotent with the immutable preload. Offline normal-child and foreign-object negatives verify single close/page/frame/disconnect events independently of Module filename. This repair has not been tested in another Chrome or original.

Closure and validation

Normal operator quit closed original Chrome985741 with actual exit0/signal-null at19:32:01.894Z; Node985586 exited0. No deliberate kill, foreign signal, retry or second original occurred. local-closure.json verifies actual PID/start identities, both PIDs and process group absent, HTTP37827 and DevTools refused, and flock free. The explicit local heavy release was posted19:32:34Z for GLTron. Box ID/state/expiry are NA; the coordinator retains future box/shutdown ownership.

Final validation:34 local tests (historical CDP-failure SIGKILL fixture excluded), 95 integration tests, actual queued producer and mid-poll negatives in both modules, standalone observer full-memory/register/flags/commands/trap parity in both modules, and the locked full build. The initial local suite also reran the existing owned Node SIGKILL-mechanism fixture; neither actual Chrome received a signal.

RUN holds exact Git SHAs, relevant binary diffs and patches for owned new files and small run-specific inputs. Modules/media remain existing-path/hash references; no source snapshot, archive, media/module copy, native install, guest-state/code/ configuration patch, box operation, npm installation or subworker was used.