MDK2 recorded lifecycle continuation: diagnostic stopped before original

The single authorized Chrome-only diagnostic failed before Puppeteer returned a Browser. The prospective recording identifies this diagnostic's actual SIGKILL caller. No original MDK2 continuation was launched, and no Chrome retry was made. Chrome894623's historical SIGKILL sender remains UNPROVED; the new evidence does not establish its cause retroactively.

Evidence is scratch/runs/20261104-mdk2-recorded-lifecycle-continuation/ (RUN). The accepted input is6f1b503a0fc278ee3aff637c15ae271318f79e22. RUN's prior-seal-verification.json verifies all503 prior sealed files,1034614 bytes, unchanged. Historical helpers and the tested mid-poll observer repair are unchanged.

The one diagnostic

GLTron's explicit19:05:41 heavy release and Stronghold's18:58:20 release were observed before acquiring /tmp/wine-local-runtime.lock. The diagnostic used the prepared, owner-filtered passive lifecycle-preload.cjs from the accepted run, the real installed Puppeteer25.7 packages, and /opt/google/chrome/chrome. diagnostic-contract.json binds its host/boot/PID/start identity, fresh nonce, candidate hashes, direct run directory and30-second browser/60-second cleanup bounds. diagnostic-packages.json hashes the installed package entries, metadata, launcher/process implementation and actual Chrome executable; no package changed.

Node957902 (start ticks186636430) spawned Chrome957916 (start ticks186636495) at19:05:55.714Z. Both are on the recorded host and boot. Launch options and the observed actual command line are retained separately. The caller selected a 15-second launch timeout within the30-second diagnostic budget. It rejected at 19:06:10.719Z, before returning a Browser. Chrome printed its own loopback DevTools endpoint at19:06:10.929Z,210ms after rejection.

The installed BrowserLauncher.launch catch invoked its close callback without a CDP connection. Its closeBrowser path waited five seconds, then called the installed Process.close. At19:06:15.722Z, the passive recorder captured process.kill(-957916, 'SIGKILL') and the actual Process.close caller stack. Chrome exited with code null/signal SIGKILL at19:06:15.736Z; its close event and Node exit1 followed at19:06:15.737Z. Browser lifetime was20022ms. This is an abnormal outcome, despite being within the time bounds.

The diagnostic caller's finally block had no Browser handle and did nothing. An additional normal-close recovery program found Chrome already absent and recorded ENOENT; it did not launch or signal anything. Ordinary page work, page/frame teardown, root disconnect and Browser.close were never reached. Their absence is explicit, so the required successful lifecycle diagnostic gate did not pass. No full original FS/HTTP prerequisite run or original release was issued. No guest/game/media, measured input, Level/Lua construction, or gameplay claim follows from this run.

local-closure.json verifies both PIDs and the Chrome process group absent, the DevTools port42809 refusing connections, and the runtime flock available. There was no HTTP server. The explicit local heavy release was posted19:11:19Z. Box ID/state/expiry are NA; the coordinator retains future box/shutdown ownership.

Evidenced caller repair

tools/mdk2-local-runtime.cjs now exposes the installed run's passive recorder and a separate active closeUnreturnedLaunch caller helper. The recorder also retains the exact observed spawn executable/arguments and bounded stderr. Recording itself still never closes, launches or signals a process.

On a rejected launch with no Browser handle, the new cleanup helper requires exactly one recorded owned child, verifies its live PID/start/host/boot identity, and waits within the explicit cleanup deadline for that child's single loopback DevTools endpoint. It connects once, observes pages/disconnect and Browser.close, requests normal closure, and requires actual exit0/signal null. Foreign or ambiguous endpoints, replaced identities, expiry, duplicate cleanup and an already abnormal exit reject. It does not suppress installed fallback signals, retry launch, or relabel abnormal closure.

The run-specific diagnostic caller is repaired to use the remaining30-second launch budget and invoke this helper immediately after a rejected launch, with at most60 seconds of cleanup. Its initial and repaired new-file patches are retained. This repaired caller has not been run against another Chrome. The original release/expiry/observer/queue/module/input guards are unchanged.

The regression uses the actual installed BrowserLauncher.launch, its real endpoint-wait timeout producer and close path, and an owned Node child that prints a delayed endpoint. The old no-Browser cleanup leaves that child live. The repaired helper consumes the actual recorded late endpoint and closes the same child through a fixture connection; its child exits0 without signals, before the installed five-second fallback. The connection is a fixture, not a claim of successful Chrome recovery. Identity, endpoint, deadline, abnormal outcome and duplicate-call negatives pass. No second browser was used to test it.

Validation:32 local tests,94 integration tests, standalone observer full-memory, register/flags/command/trap parity in both modules, accepted111-helper/four-Git dependency/both-module guards, and the locked full build. The already accepted CDP-reject SIGKILL fixture was excluded from this rerun; its historical proof is retained, while the new fixtures request only normal Node-child exit.

RUN contains exact Git SHAs, relevant binary diffs and new-file patches. Existing module/media locations remain references; no source archive, media copy, module copy, npm installation, box operation or subworker was used.

Next is a separately authorized fresh Chrome-only diagnostic using the repaired caller. Its normal, fully recorded closure is still required before considering another original construction attempt. The accepted mid-poll suffix retains an UNKNOWN prefix; missing demand/finalizer/buffer history must still stop any future construction observation.