MDK2 poll-history repair and unresolved Chrome termination

The observed poll-history failure has a tested observer repair. Chrome894623's termination cause remains unproved, so this assignment launches no new original continuation. The two failures remain distinct. This is an offline repair and diagnostic handoff, with no gameplay claim.

Evidence: scratch/runs/20261104-mdk2-poll-history-lifecycle-repair/ (RUN). The prior handoff is mdk2-startup-timing-onward-20261104.md, published in 7bd2cd305eca803c3463ac0bd10fc174b8a4e9e8. RUN/analyze-prior.cjs verifies every one of its1364 sealed files,16597901 uncompressed bytes, including retained gzip forms. Its seal SHA256 is 688091c2ee3d14904b7336e6110b91f1eadc6ca2e0ad9e2cf26ac71f00b92546. Historical evidence/helpers are untouched; initial failure logs are retained.

Actual attachment ordering

The original owning Worker was slot0/endpoint1/thread1, with file Worker slot1/endpoint2/thread2. The full frozen main.link identity, lifecycle and queue barriers pass before and after command101. The canonical Module binding and installed-before-instantiation audit pass. Installation creates a dormant wrapper; the observer and native tracing start only at the construction command. Those are different times, as the installed source explicitly specifies.

At attachment, step3312, main EIP is4515cd and ESP074ffc24. Its first four stack words are4522d7,10,44fb84,005e3840; the last is the actual level1 request. The resource is active, type10, incomplete, size192320. The original instructions in RUN/original-instructions.txt establish this exact chain:

PC Original instruction / observation
44fb7e /44fb7f push request; call4522d0
4522d0 /4522d2 push10; call4515c0
4515c0–4515c7 call SleepEx with the original arguments
4515cd ret; this is the attachment point and first delivered boundary
4522d7 read completion; second boundary, old observer rejects here
4522dc /4522df add esp,4; ret to44fb84

The first two boundary addresses are reproduced by the actual native producer; the historical report itself retained the count2 and failure, rather than a complete raw boundary list. The sealed initial registers/stack, original decoded instructions and queued-producer reproduction establish the ordering. We do not claim a historical raw trace that was never recorded.

This is attachment in the middle of a native poll, with its prefix unobserved. It is not evidence of a wrong hook/owner/correlation or an unsupported x86 instruction ordering. Command101 requested and ran16 slices, reaching3328, before its assertion rejected. Missing poll history was UNKNOWN, never a guest stall. Frozen file output advanced8192→61440 bytes, compressed read2048→8192; completion remained0 and no Level table or original gameplay was established.

Repair and rejection boundaries

The observer can now recognize only this exact initial Sleep-return suffix. It additionally pins the Sleep wrapper and original request call bytes. Its first sample must show the original return address,10ms argument,44fb84 caller, and the same active type10 request. The first native boundary must match that sample. The next must be4522d7 with ESP advanced4 and the same remaining stack; the following must be44fb84 with ESP advanced12, the same request and the actual completion return value. A mismatch is terminal.

It emits poll-prefix-unobserved, poll-suffix-read, and poll-suffix-end. It never creates a poll-entry record, seeds poll, emits poll-return for the partial call, or promotes unknown history into known history. The original poll history missing assertion remains for every unqualified completion-read boundary. Future fully observed polls use the existing entry/return checks. Deadline, record/boundary bounds, missing history, foreign instance/thread, loader, bytes, request, completion/finalizer and script-buffer checks remain. This repairs the evidenced poll suffix, not every possible attachment scope. An earlier demand-load call is also outside this observed interval; no history for it is invented. A later missing demand/finalizer/buffer history still stops observation. No complete resource-to-Lua sequence is established by this repair.

Both retained canonical/companion modules execute the actual original RET/MOV/ADD/RET sequence in the pinned shipped Worker through its ordinary queue and exact composed preload. The old observer fails after2 boundaries; the repair passes, with completion0 throughout. Full-memory hashes and registers agree at the identical poll endpoint; guest-image memory, registers and command sequences also agree after delivering the caller boundary. A whole-memory comparison after that extra caller slice was initially attempted and is preserved as a failed test: the old rejection had already disabled tracing, so that comparison does not have identical trace configurations. Existing complete normal/observer full-memory parity guards still pass in both modules.

Bad initial stack, stale stack after sampling, changed request and skipped initial boundary reject. Existing generic missing-history and all real composed Worker lifecycle/owner negatives still reject. The fixture executes no game and creates no Lua state or callbacks. Integration keeps the original de6 observer pin and adds the exact repaired observer SHA256 341f705e65d292ab014146af0594cc0531913428f14b0f47a125f7fba4acf0cb. All111 helper/four Git dependency guards and432 actual-filesystem/full-HTTP pins remain. No historical helper is patched. The standalone observer test now reads compressed historical seal/inventory inputs without restoring or overwriting them; its native qualification fixture remains byte-identical to de6.

Chrome termination: mechanism available, attribution unknown

Command101 failed18:23:32.831Z. Shot102 failed with detached Frame at18:24:42.934Z, 70.103 seconds later. Quit103 began18:24:59.485Z and cleanup recorded SIGKILL, exit null, at18:24:59.490Z:5ms later. The detached-frame failure preceded quit by16.551 seconds. These timestamps are observations/cleanup records, not the time Chrome actually received SIGKILL. Quit ACK is operator handback, not proof of orderly Chrome termination. Prior abnormal outcome remains unchanged.

The installed Puppeteer25.7.0 BrowserLauncher.closeBrowser tries CDP Browser.close. On rejection it calls the installed browser-process close(); on Linux, a still live process is killed using process.kill(-pid,'SIGKILL'), with a child.kill fallback. Driver exit/SIGINT can also reach kill; SIGTERM/SIGHUP reach close. The original driver calls browser.close during handback. Trace failure itself stops tracing and saves the failed ACK; it does not call browser.close. The 40-second command timeout,60-second bootstrap limit and absolute browser timer do not match the recorded successful bootstrap/short commands and early quit.

The installed launcher and process implementation were exercised with owned Node children: graceful CDP completion exits0; rejected CDP close invokes the actual Linux process-group SIGKILL path, with a captured stack through BrowserLauncher.closeBrowser. RUN/lifecycle-mechanisms.json records both cases, timestamps, actual PID/start identities, exact installed source paths/hashes and kill stacks. These fixtures prove a mechanism exists; they do not identify the sender or trigger of the historical Chrome signal. No OOM inference or rejected security diagnostic is used. These are the currently installed package bytes. The prior run did not seal Puppeteer package hashes, so their exact identity with that process's dependency bytes is not cryptographically established either.

Historical records lack an exit-event timestamp, root browser disconnect time, page/target teardown sequence, close-request result and local signal-call stack. The early detached Frame alone does not prove root browser death. A close fallback during quit is compatible with the5ms cleanup interval, but remains an inference, not established attribution. External signal attribution is also unknown. No source-only inspection can recover those missing events.

Executable diagnostic/recovery recipe

Offline reproduction, without Chrome or a game:

RUN=/home/user/wine-assembly/scratch/runs/20261104-mdk2-poll-history-lifecycle-repair
node "$RUN/analyze-prior.cjs"
MDK2_TEST_EVIDENCE="$RUN" node --test --test-name-pattern=mid-poll test/test-mdk2-construction-driver-integration.js
MDK2_LIFECYCLE_RESULT="$RUN/lifecycle-recheck.json" node --test --test-name-pattern='installed Puppeteer close path' test/test-mdk2-local-runtime.js

The new exported local.traceOwnedLifecycle records actual child spawn identity, exit/close, bounded stderr, root disconnect, page errors/close/frame detach, browser.close calls/results and owned process.kill/child.kill callers. It never launches, closes or signals anything. Its profile/executable filter excludes other launches. Test fixtures verify the actual SIGKILL caller is captured, foreign browser admission and overwriting an existing log reject, and normal closure stays distinct. The helper is not automatically installed by this assignment and does not renew the prior release or grant a launch.

RUN/lifecycle-preload.cjs is a prepared diagnostic preload for a separately authorized fresh owned diagnostic driver. It intercepts only that direct run's CommonJS Puppeteer require, preserves launch arguments/results, and attaches the recorder before spawn. It watches browser/pages after they are created. Its no-launch smoke and actual run-local require/foreign-profile rejection are tested; the initial immutable-ESM-export failure log is retained. For a newly authorized diagnostic run with its own finite contract and launch script:

MDK2_LIFECYCLE_RUN="$FRESH_DIAGNOSTIC_RUN" node \
  --require "$RUN/lifecycle-preload.cjs" \
  "$FRESH_DIAGNOSTIC_RUN/diagnostic-launch.cjs"

This is a handoff recipe, not an executed or authorized new game continuation. Use a Chrome-only diagnostic first, retaining lifecycle/stderr and close failure logs. If a future original attempt is separately justified/authorized, preserve all original view/input/owner/module/HTTP guards and install this recording before launch; it cannot retroactively establish the old signal's cause. No fresh original should be spent merely reproducing a known fault. Do not rerun the old release, edit ACKs or sealed helpers, force Level/Lua/native installation, relax guards, infer gameplay from file progress, or call SIGKILL normal closure.

Both requested resource releases were observed: GLTron18:36:33Z and Stronghold route handoff18:39:17Z. This assignment did not start Chrome, a Node HTTP server, or a guest, and did not reserve the runtime flock. The old Node894431 and Chrome894623 remain absent; HTTP33021 refused and the runtime lock was available at final verification. Box ID/state/expiry: NA; coordinator owns future boxes and shutdown. No boat/API/SSH-box/npm/allocation/subworkers were used.

Final validation:31 local tests,94 integration tests, standalone observer normal/observed full-memory/register/flags/trap parity in both modules, and the locked full build. Evidence includes exact Git SHA, relevant binary diff and new note patch; media/modules are existing path/SHA references, with no source archives or media/module copies. Next: resolve the lifecycle attribution gap with prospective diagnostic evidence before considering another original run.