MDK2 Medium release: admitted rest, queued native keyup

Input is 87a175b1f39159ed94b8d789d8bfc8346c2f3b1e and mdk2-honest-menu-medium-gameplay-20261104.md. Evidence is scratch/runs/20261104-mdk2-medium-release-gameplay/ (RUN). The single authorized original attempt stopped at a concrete selector budget fault. Medium's keyup remained queued after 464 rest steps. Native Medium release consumption/callback, loading and ordinary gameplay remain unproved. No second actual attempt occurred, and no additional causal gameplay repair was established offline.

The local contract now names this assignment and requires Stronghold's exact stronghold-named-observation-gameplay-nov04 heavy release at or after 22:00:55Z and its file release at or after 22:05:30Z, following that actor's latest ownership entry. Missing, stale, foreign, quoted and superseded entries are rejected. The actual launch used the real board entries. The fixed 32-step Medium tail repair is unchanged: 512 press steps, 544 aggregate steps, one shot after acknowledged balanced keyup. The 113-command/240-second menu guards and all native input/callback/ACK checks remain.

Before Chrome, the preceding attempt's exact command/native prefix replay again rejected the required rest under the old selector and submitted exactly 32 steps under the repaired selector. Press ceiling and tail exhaustion negatives passed. The generated initial/startup commands and open/New Game/ Medium review field sets traversed the original Python caller transport and ACK client offline. The fresh complete entry, using the same run-specific wrapper as the actual launch, reached an explicit plain Puppeteer mock with independent public and native child-spawn prohibition. It spawned zero Chrome, closed both HTTP servers and left both ports refusing connections. No ESM namespace exports were replaced. Validation passed 42 local checks, 144 observer/integration checks and the full locked build.

Every fresh complete capture was personally viewed: startup steps 0/512/1024/1536/2048/2560/3072/3584, then native selector captures 64 and 67. The first three images showed startup/window creation; 1536 through 3072 were black. At 3584 the complete menu already showed New Game with a red highlighted controller, without the legal overlay. A 16-step observation produced the separately viewed 3600-step image with the same visible selection. Native New Game readiness passed. No Escape input was needed or submitted. Its balanced Enter selection produced the separately viewed 3648-step difficulty dialog with orange highlighted Medium. Medium's own visual fields, preapproval and native readiness checks passed.

Medium's press became native-positive after one 16-step quantum. Its keyup, command 79, was acknowledged and balanced the trusted DOM history. The first 32-step rest was submitted and completed. Subsequent 16-step rest quanta also completed, but the original queued WM_KEYUP remained present through step 4128: 464 rest steps in total. Enter raw/previous remained 1/1, active 0; menu.current remained diff, showmenu 1, freeze nil and scheme menu. Those readbacks establish neither release consumption nor the Medium callback. The selector ledger reached 544 steps and rejected the next quantum before submission. This was the first actual fault. All 137 submitted commands, including normal quit, had successful ACKs; every submitted step completed.

The native observer retained 29 checkpoints, 577000 boundaries and 31 sample records under one unchanged deadline. Its owning Worker terminal report has no fatal error and status terminal-cleanup. There are no demand, finalizer, consume, poll-return, script-buffer or new LuaLevel construction records. The fixed 50-million-boundary/360-second/512-record contract and byte, nesting, owner, module, thread, input, frozen-state and ACK guards were not changed or reset. Original guest instructions, configuration and 16ms pacing were not written.

The exact new command/native prefix was replayed offline against the repaired selector using explicit plain fixture inspection bytes, without reading or rehearsing original game images. It reproduces admission of the first 32-step rest, every command through 135, the pending native keyup and the 544-step rejection before another submission. This validates the blocker and the original repair's limited effect; it does not supply a later callback. RUN contains offline-first-fault-replay.json and actual-first-fault-analysis.json with these distinctions explicit.

Stronghold's separate worker identified and tested default log_block RPC overhead in the shared bridge. That change is not part of MDK2's immutable runtime pins. It is a candidate for further offline diagnosis and exact dependency integration, not proof of the cause or a repair of this run's queued keyup. Further diagnosis must demonstrate release/callback progress under the existing finite guards before another separately authorized original attempt. Expanding the selector or native observation limits was not attempted.

Normal quit command 136 closed Chrome PID 1268267/start ticks 187771286, actual exit 0/null at 22:19:51.000Z. Node PID 1268119 and external flock PID 1268118, both start ticks 187771001, exited 0 and were absent. HTTP ports 36085/34901 and DevTools 41265 refused connections; the exact profile was removed after process absence and the external flock was verified free at 22:20:10Z. Assignment-bound LOCAL HEAVY RELEASE was posted at 22:20:20Z. Box ID, state and expiry are NA; coordinator retains shutdown ownership.

RUN records launch Git, relevant binary diff, final binary diff and this new note as a patch. Actual original runtime Git remains 1329b1ba6eb00144ff77d4cd92a8513d4fc953ad, with modules cffca869b541e6eff12ae3a6b74c4476977e0a7e23cd43d8ec5ba8faecd4bcee and 580f49f3c78d26e6855efa0beb37046d03ea24690ff98cfcb3e94ce5e6cb1ee3. Needed released MDK JSON/log inputs were restored with gzip -dk using released-headroom-2209.jsonl, verifying each original SHA; the mapping is retained in restored-inputs.json. Existing media/modules are referenced by path and SHA. No source snapshots, tars, media/module copies, npm installation, boats or subworkers were used. This assignment stops after publication and path release, with a concrete unresolved native-input blocker.